£815

 

LIBRARY
Michigan State
University

 

 

 

This is to certify that the
thesis entitled

EXPLORATION AND EXPLOITATION:
A NEW EXPLANATION OF DIFFERENTIAL
GOAL-SETTING EFFECTS

presented by

PAUL G. CURRAN

has been accepted towards fulfillment
of the requirements for the

degree In PSYCHOLOGY

 

 

%/M

Major Professor’ s Signature

/,2//9’/0<?

Date

MSU is an Affirmative Action/Equal Opportunity Employer

 

 

PLACE IN RETURN BOX to remove this checkout from your record.
TO AVOID FINES return on or before date due.
MAY BE RECALLED with earlier due date if requested.

 

DATE DUE DATE DUE DATE DUE

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

5/08 K:/Prolecc&Pres/ClRC/DateDue.indd

 

EXPLORATION AND EXPLOITATION:
A NEW EXPLANATION OF DIFFERENTIAL GOAL-SETTING EFFECTS

By

Paul G. Curran

A THESIS

Submitted to
Michigan State University
in partial fulfillment of the requirements
for the degree of

MASTER OF ARTS
Psychology

2008

ABSTRACT

EXPLORATION AND EXPLOITATION:
A NEW EXPLANATION OF DIFFERENTIAL GOAL-SETTING EFFECTS

By
Paul G. Curran
Goal-setting theory has been shown over the years to have a number of

experimental outcomes inconsistent with previous theory. An experiment was designed
in order to attempt to intentionally produce these outcomes in a controlled design, such
that explanation could be provided a priori as opposed to post hoc. It is suggested that
these results, as well as more standard results, could be explained using the constructs of
exploration and exploitation. Results basically supported main predictions concerning

these ideas.

TABLE OF CONTENTS

LIST OF TABLES ............................................................ v
LIST OF FIGURES ........................................................... vi
INTRODUCTION ............................................................... 1
Complexity .............................................................. 5
Exploration and Exploitation ........................................ 10
Exploration and Problem Representation .......................... 12
Exploration/Exploitation and Goal-Setting Effects ............... l4
Exploration and Reinforcement Learning ......................... 18
Summary and Hypothesis ............................................ 20
Individual Difference Effects ....................................... 21
Curiosity ...................................................... 22
Goal Orientation ............................................. 23
METHOD AND RESULTS .................................................. 25
Study 1: Pilot .......................................................... 25
Participants ................................................... 25
Procedure ...................................................... 25
Results and Discussion ...................................... 27
Study 2: Pilot 2 ........................................................ 29
Participants ................................................... 29
Design ......................................................... 30
Procedure ...................................................... 31
Measures ...................................................... 32
Perceptual complexity .............................. 32
Perceptual difficulty ................................. 32
Results ......................................................... 32
Study 3: Test of Goal-Setting Effects .............................. 34
Participants .................................................... 34
Design .......................................................... 35
Procedure ...................................................... 35
Measures ....................................................... 38
Trait level goal orientation ......................... 38
Trait level curiosity ................................. 38
Perceptual complexity .............................. 39
Perceptual difficulty ................................. 39
Exploration (operationalization) .................. 39
Results ......................................................... 40
Manipulation check ................................. 40
Main Analyses ....................................... 40
Discussion .............................................................. 46
Implications ................................................... 50

iii

Complexity ........................................... 50

Exploration and problem representation ......... 51

Exploration and goal setting theory ............... 51

Exploration and reinforcement learning .......... 52

Other implications ................................... 53

Practical implications ............................... 54

Limitations ..................................................... 54
Next steps ...................................................... 56
Conclusions .................................................... 57
APPENDICES .......................................................... 59
REFERENCES ......................................................... 71

iv

LIST OF TABLES
Table 1: Regression of cumulative exploration on performance over time by world. ....59

Table 2: Linear regressions of proposed moderators on exploration: results for
hypothesis 3, 4a, and 4b (EXl) .................................................................. 60

Table 3: Linear regressions of proposed moderators on exploration: results for
hypothesis 3, 4a, and 4b (EX2) .................................................................. 61

Table 4: Results of repeated measures analysis on performance and exploration over
time by goal condition and world ................................................................ 62

LIST OF FIGURES

Figure 1: ‘Do your best’ goals .......................................................... 63
Figure 2: Specific, difficult goals ........................................................ 64
Figure 3: Performance over time ......................................................... 65
Figure 4: World 1 .......................................................................... 66
Figure 5: World 2 .......................................................................... 67
Figure 6: World 3 .......................................................................... 68
Figure 7: Start of trial (zoomed out) .................................................... 69
Figure 8: Afier some exploring in trial (zoomed out) ................................. 70

Images in this thesis/dissertation are presented in color.

vi

INTRODUCTION

The finding that specific, difficult goals improve performance relative to ‘do your
best’ goals has been demonstrated in thousands of studies throughout the course of nearly
40 years (Locke, 1968; Locke & Latham, 1990; Locke & Latham, 2002; Locke, Shaw,
Saari, & Latham, 1981). Beyond this finding, research has also addressed how goals
produce such a shift in performance. Goals are said to accomplish this shift through four
distinct mechanisms: goals “direct attention and effort toward goal-relevant activities and
away from goal-irrelevant activities”, “have an energizing function”, “prolong effort” via
persistence, and lead “to the arousal, discovery, and/or use of task-relevant knowledge”
(Locke & Latham, 2002, p706-707). In short, goals increase and direct personal
resources to create a boost in performance.

Despite such findings, goals do not produce a strong goal setting effect in every
situation. There exist situations in which specific difficult goals produce a much weaker
goal setting effect over do your best goals, or even a negative goal setting effect where
those given ‘do your best’ goals outperform those given specific, difficult goals (Locke,
Shaw, Saari, & Latham, 1981). For instance, in a study by Kanfer and Ackerman (1989),
participants in a specific, difficult goal condition were outperformed early in the task by
those in a ‘do your best’ condition. Huber (1985) demonstrated the same effect on a
simple maze task with both easy and difficult versions of the maze. Sweller and Levine
(1982) found that the presence of a goal (relative to a no-goal condition) led to such a
detriment in task performance that a simple maze task was rendered virtually unsolvable
in the goal condition. Earley, Connolly, and Lee (1989), as well as DeShon and

Alexander (1996) showed a negative goal effect on similar, moderately complex,

versions of a multiple cue probability learning task. Burns and Vollmeyer (2002) also
found a negative goal setting effect on transfer performance on a multiple cue probability
learning task. While these studies do not reach the number of studies showing a positive
goal setting effect they are still too numerous to dismiss as experimental flukes. What,
then, is occurring?

Early speculation around this idea led to the suggestion that task complexity
might account for weaker cases of the goal setting effect (Locke et a1, 1981). Using
Wood's (1986) model of complexity, Wood, Mento, and Locke (1987) performed a meta-
analysis of the goal setting literature and showed that, in general, as task complexity
increased the positive goal setting effect decreased. Thus, simple tasks showed the
strongest positive goal setting effect, and more complex tasks showed the weakest. This
finding has been explained in numerous ways, usually focusing on aspects of cognitive
resources such as attention and effort (Kanfer & Ackerman, 1989; Locke et a], 1981).

While Wood, Mento, and Locke’s (1987) explanation of task complexity
moderating the goal setting effect does seem to work for a large number of studies there
are three experimental results that are contradictory. First, the task complexity
explanation cannot account for studies with simple tasks where specific, difficult goals
show diminished performance in relation to those with ‘do your best’ goals. A number of
these studies have been highlighted above (Burns & Vollmeyer, 2002; Huber, 1985;
Sweller & Levine, 1982). Second, the task complexity explanation cannot account for
studies with complex tasks where specific difficult goals show a strong goal setting effect
- that is, notable increased performance over subjects given ‘do your best’ goals. While

not as numerous, these effects do exist (Ivancevich, 1974; Pritchard, Jones, Roth,

Stuebing, & Ekeberg, 1988). Finally, the task complexity explanation does not account
for situations in which a single task can produce vastly different goal setting effects
through non-complexity based manipulation. While no studies have directly addressed
this issue, results such as those of Kanfer and Ackerman (1989) can be interpreted from
this framework and will be discussed later in this paper. Nevertheless, numerous studies
have demonstrated the first two experimental situations. Given these findings, goals must
be acting through a means other than task complexity.

Wood (1985) suggested that specific, difficult goals on complex tasks might lead
to suboptimal search processes. Others, such as Barley, Connolly, and Lee (1989),
suggested that specific, difficult goals might lead to excessive strategy search. Seijts and
Latham (2001) suggested that performance (score) goals and learning (exploration) goals
might have differential effects. These different explorations, at their roots, lead back to a
basic psychological distinction between exploration and exploitation.

Exploration has been treated in many different ways in the various fields of
psychology, but it may in fact be a simple dictionary definition that best encapsulates it —-
“the investigation of unknown areas” (Random House, 2006). When studied, exploration
is often defined and studied in relation to its counterpart of exploitation. Exploitation is
the behavior of acting on a system in the best known way. Exploitation comes at the cost
of exploration the same that exploration comes at the cost of exploitation. While
exploration is often said to be driven by motivations (such as an animal searching for
food to survive), exploration is potentially motivated by different mechanisms (V 055 &
Keller, 1976); situations “for which there appears to be no goal object or condition to and

for which the organism responds” (Fowler, 1965, p23). In terms of goal setting,

exploration can be defined as actively sacrificing known optimality toward a goal in
order to search the unknown.

The purpose of this research is to suggest and show that the way in which specific
difficult goals may be causing both positive and negative goal setting effects is through
their effect on exploratory behavior. Specific goals, according to goal setting theory,
“direct attention and effort toward goal-relevant activities and away from goal-irrelevant
activities” (Locke & Latham, 2002, p706). Exploitation — action on the task toward the
goal — is a goal-relevant activity and produces positive goal effects in thousands of
research studies. It involves striving toward the goal in the best way one knows how. In
this case, exploration for the sake of exploring is goal-irrelevant. Put simply, it is
suggested that specific goals limit exploration relative to ‘do your best’ goals.

It is thus further suggested that this effect on exploratory behavior produces
different goal effects due to the fact that performance on some tasks requires more
exploration than others. If a task requires exploration to perform well (Burns &
Vollmeyer, 2002; Sweller & Levine, 1982), specific goals that inhibit this exploration
will decrease performance and produce a negative goal effect (Burns & Vollmeyer, 2002;
Sweller & Levine, 1982). If exploration is not needed to perform well on a task (Latham
& Kinne, 1974; Locke, 1968) then specific goals will help focus effort on the appropriate
method of completing the task and produce a positive goal effect (Latham & Kinne,
1974; Locke, 1968).

This explanation also holds potential to more deeply explain the results of Wood
et a1 (1987). Tasks that are generally more complex are likely to require more

exploration to complete (producing a negative goal effect by limiting exploration) while

simpler tasks are likely to require less exploration to complete (producing a positive goal
effect by limiting exploration). Thus, while complexity can appear to act as a moderator,
it is suggested that what is occurring at a more basic level is the underlying effect of
specific goals on task exploration.

To this end, complexity will first be examined before moving on to exploration
and exploitation. With this foundation an experiment will be proposed. Proposed
individual difference moderators of effects will also be discussed.

Complexity

Before one can reasonably assess the validity of the task complexity moderation
on goal setting effects, one must address a number of questions, such as ‘what is
complexity?’ Much of the work on complexity is based on the framework laid out by
Wood (1986). This fiamework is that on which the study by Wood, Mento, and Locke
(1987) was carried out. It is centered on three aspects of complexity: component
complexity, coordinate complexity, and dynamic complexity. In short, Wood suggests
that these three aspects of complexity interact to determine how complex a task is.

In order to grasp Wood’s definition of overall complexity one must thus review
Wood’s definitions of the aspects of complexity:

“Component complexity of a task is a direct function of the number of distinct

acts that need to be executed in the performance of the task and the number of

distinct information cues that must be processed in the performance of those acts”

(Wood, 1986, p66). “Coordinative complexity refers to the nature of the

relationships between task inputs and task products. . . [including] timing,

frequency, intensity, and location requirements for performances of required acts”

(Wood, 1986, p68). “Dynamic complexity is due to changes in the states of the

world which have an effect on the relationships between task inputs and products”
(Wood, 1986, p71).

Following from these definitions, Wood defines total task complexity as “a
function of the component, coordinative, and dynamic complexities of the task” (Wood,
1986, p74). Thus, a clear and concrete definition of complexity at the general construct
level remains unsupplied by Wood (1986). Moreover, a competing and no less valid
typology of complexity was produced by Campbell (1988) just two years later and
supplies a different view on complexity in which complexity is comprised of four
dimensions different than Wood’s. Campbell (1988) claims that the four important
dimensions of complexity are multiple paths, multiple outcomes, a conflicting
interdependence among paths, and the presence of uncertain or probabilistic linkages.
While there is slight overlap with Wood’s (1986) dimensions, this produces a much
different picture of what complexity is and highlights the fact that complexity as a
singular construct is incredibly difficult to define. Even a search in a dictionary produces
a fruitless and circular result: “Complexity: 1. The quality or condition of being
complex. 2. Something complex: a maze of bureaucratic and legalistic complexities”
(American Heritage Dictionary, 2006).

Further, while component complexity is fairly concrete and coordinative
complexity somewhere between concrete and abstract, dynamic complexity is placed as
the most important dimension of task complexity while simultaneously possessing the
highest level of abstraction. Transfer tasks are a useful example of this, and Wood’s take
on dynamic complexity begs the question of how they might be rated on this scale. Is a
task complex because of its characteristics, or because of how it is administered? If a
transfer task is presented in which task rules change slightly, can it possess its own rating

of complexity or must now the entire task and transfer task be rated with the same

values? Does the task itself possess complexity outside of the situational boundaries in
which it is presented? The answers to these questions are well beyond the scope of this
paper, but highlight a further difficulty in the conceptualization of complexity as a whole.

With such difficulty in conceptualization of complexity it is no surprise that
operationalization of complexity proves to be equally difficult. Wood (1986) puts forth a
series of concrete and quantifiable equations that can be used to calculate the values for
the different dimensions of complexity, and in fact even presents two example cases in
which he presents explanations and values for these figures. Despite this, however,
Wood is forced to admit that “at this point, we cannot specify the exact form of the
relationship between the different types of complexity and total task complexity” (Wood,
1986, p74). Without this relationship, Wood’s final equation of total task complexity
(total complexity = a*component complexity + B*coordinative complexity + y*dynamic
complexity) possesses three unknowns — the different dimensions’ weighting variables a,
13, and 7. Thus, while two tasks that possess values in only one of the three dimensions
can be compared in terms of complexity, comparing values across the different
dimensions is functionally equivalent to comparing apples and oranges.

Wood, Mento, and Locke (1987) further state that “task complexity is ordinal in
nature” (p417), further distancing the field from Wood’s implication that complexity may
be represented on a mathematically defined ratio scale. This is carried out in their
operationalization of complexity, as even they (Wood, Mento, Locke, 1987) do not use
the mathematical formulas set forth a year earlier by Wood (1986) and instead use
aggregate ratings by coders on a ten-point scale. This coding on 125 tasks produces a

range of complexity ratings from only one to seven, with one representing simple

reaction time tasks and seven representing “science and engineering” (Wood, Mento,
Locke, 1987, p 418). Over 60 percent of the tasks are represented by ratings of two or
three on the ten point scale. Beyond these problems, but possibly underlying them, is the
fact that - regardless of the simplifying assumption that task complexity can be
represented on an ordinal scale - Wood, Mento, and Locke (1987) have still not overcome
the pitfall put forth by Wood (1986): “we cannot specify the exact form of the
relationship between the different types of complexity and total task complexity” (Wood,
1986, p74). Thus, total task complexity as laid out by Wood cannot be represented even
on an ordinal scale, by current measurement methods. Only effects can be measured,
tasks that differ on different types of complexity have the potential to look the same on
ratings. The fundamental assumptions underlying the operationalization in Wood,
Mento, and Locke (1987) were shown to be unreachable a year prior by Wood (1986)
himself.

As already discussed, complexity is suggested to be linked to the goal setting
effect in such a way that larger positive goal setting effects are found on simpler (less
complex) tasks while smaller positive goal setting effects will be found on more complex
tasks. Wood, Mento, and Locke (1987) showed that this relationship holds for the most
part using the framework of task complexity put forth by Wood (1986) and discussed
above. The implication follows that goal-setting effects should not be prone to
manipulation if such manipulation is not done on the complexity of the task.

Even accepting that in largely distinct tasks (very complex and minimally
complex) complexity can be measured ordinally, data presented earlier in this paper still

falls counter to the hypothesis of Wood, Mento, and Locke (1987) that less complex tasks

should possess more powerful positive goal setting effects than more complex tasks.
There exist tasks that possess relative low levels of complexity (maze tasks, simplified
MCPL tasks) in which participants still experience a negative goal setting effect (Burns
& Vollmeyer, 2002; Huber, 1985; Sweller & Levine, 1982). There exist tasks that
possess relative high levels of complexity (complex jobs in workforce, management) and
show large positive goal setting effects (Ivancevich, 1974; Pritchard, Jones, Roth,
Stuebing, & Ekeberg, 1988). Finally, Kanfer and Ackerman (1989) provide an accidental
demonstration of the third situation postulated to show a failing of the theory of
complexity as a moderator of the goal effect: a situation in which a single task produced
different goal setting effects through non-complexity based manipulation. This effect
will be discussed later in the paper.

A task, according to Wood (1986), and by extension Wood, Mento, and Locke
(1987), is complex because of the three dimensions of complexity and these three ideas
alone. Goals, unless they alter one of these three components, cannot make a task more
or less complex. In Kanfer and Ackerman (1989), goals possess no complexity altering
qualities. In this experiment, it is first shown that when participants are given goals at the
start of the experiment it produces a negative goal setting effect — specific, difficult goal
participants are outperformed by those in a ‘do your best’ goal condition. When this
same goal manipulation is given later in the experiment (all participants receive ‘do your
best’ goals in trials one to four, specific, difficult goals begin in goal condition in trial
five), the goal setting effect is reversed and becomes a positive goal setting effect in
which do your best goal participants are outperformed by those in the difficult specific

goal condition. In Wood’s (1986) framework there is no possible way this manipulation

could have changed the complexity of the task, yet the goal-setting effect was changed
nonetheless. From an objective task complexity standpoint this is impossible.
Cumulative with the previous evidence, Kanfer and Ackerman (1989) show that
examples exist to counter all three of the main implications of Wood, Mento, and Locke’s
(1987) finding. What, then, is actually the cause?
Exploration and Exploitation

The psychological questions underlying the topic of exploration have been around
nearly as long as the science of psychology. In attempts to explain behavior, early
psychologists reasoned that just as instincts drive other animals, so too must human
behavior be driven by instinct (James, 1890). This idea eventually evolved into the
notion that motivational drive states, such as hunger, thirst, and pain, caused behaviors
that would return the being to a state of homeostasis (Cannon, 1932; Carr, 1925). In the
case of goal setting, this would reasonable account for actions taken to move oneself
toward a goal state (i.e. exploitation). However, this idea failed to account for situations
in which an organism would investigate a puzzle without the presence of a primary drive
(Harlow, 1950). The following investigation of exploratory behavior became very much
intertwined with the idea of curiosity as a possible drive of this behavior (Fowler, 1965).
In summary of this following research it was recommended that “exploration should be
used to describe observable behavior, whereas curiosity should be used to describe the
corresponding hypothetical construct” (Voss & Keller, 1983, p150). Furthermore,
exploratory behaviors “are visual, linguistic, haptic, motoric, or intellectual activities that
make new information available to the individual” and “are used to reduce uncertainty”

(V oss & Keller, 1983, p150). This new information has only the potential of being

10

useful, however, and the time spent collecting it is time not spent exploiting the
information one already has. This balance of exploration and exploitation as a resource
allocation issue is an issue of importance.

Over time, many different areas have come to study this dilemma of exploration
versus exploitation, knowingly or not. The topics of game theory (Jervis, 1988),
evolutionary computing and genetic algorithms (De Jong & Spears, 1993), and even
dating and mate selection (Das & Kamenica, 2005) have all come to the conclusion that
there exists a balance between exploration and exploitation that must be met to obtain
optimality. One must spend time exploring and collecting information, but must also
realize when the potential for future information must be sacrificed for immediate action
with an incomplete knowledge set. This subtle balancing act is pushed to even farther
limits in fields such as foraging theory (Pyke, 1984), where neglecting exploration for
exploitation or exploitation for exploration can literally result in death at both the
individual and species level. If a herd of antelope stay at the same watering hole and
eventually eat all the vegetation around it without allowing for re-growth, they will
eventually starve. A similar herd that moves outside of their known world to find other
watering holes while using only some of the vegetation of the first will now not only have
food, but the knowledge of where to find more.

The study of exploration and exploitation has also found root in organizational
research. March (1991) noted the balance that must be kept between these two ideas, and
that deviating from optimal in both directions could be disastrous:

“Adaptive systems that engage in exploration to the exclusion of exploitation are

likely to find that they suffer the costs of experimentation without gaining many

of its benefits. . . .Conversely, systems that engage in exploitation to the exclusion
of exploration are likely to find themselves trapped in suboptimal stable

ll

equilibria. As a result, maintaining an appropriate balance between exploration

and exploitation is a primary factor in system survival and prosperity. (March,

1991, p71)”

March (1991) also takes note of the vulnerability of exploration in that it produces
less certain outcomes that only have a chance of paying off. Just as in the research
mentioned above, exploration is not a sure bet. In do or die situations exploration may
hold much more risk than benefit. Other organizational research has taken place in the
context of exploration and exploitation, from research on innovation (Nooteboom, 1999),
resource allocation (Garcia, Calantone, & Levine, 2003), and even strategic ventures and
alliances (Rothaerrnel & Deeds, 2004).

Exploration and Problem Representation

Work in the field of problem space and problem representation provides a link
between goals and the balance of exploration and exploitation. “A problem space has an
initial state, a goal state, and a set of operators that can be applied that will move the
solver from one state to another” (Dunbar, 1998, p 290). In this framework, the
representation of the goal state can fiindamentally alter the behaviors that individuals
enact to move from an initial state to the goal state. As a concrete example, Miller,
Lehman, and Koedinger (1999) used a task called electric field hockey in which
participants tried to push a puck between obstacles and into a net. Participants were
given either a specific goal, a nonspecific goal, or an appropriate specific goal. The
appropriate goal was designed to push participants toward the optimal manner of problem
completion through the accurate representation of the task. The appropriate specific goal
and the nonspecific goal groups both produced better performance than the specific goal

group, suggesting that participants could find the optimal behaviors by search and

12

exploration (no goal), or by specific direction toward it (appropriate goal). The linkage
of exploration and search to performance in the no goal group was supported by
significantly increased scores on a knowledge test relative to other conditions. In all, it
was suggested that the lowered performance in the specific goal condition was due to the
primes toward task-irrelevant actions built into the standard goals. In fact, it was
suggested that one of the benefits of the no goals condition was that “the no goal
condition was not ‘distracted’ by having to focus on irrelevant. . .relationships of standard
[electronic field hockey] play” (Miller, Lehman, & Koedinger, 1999, p320). The
appropriateness and inappropriateness of these goals can be viewed as the balance of
exploration and exploitation they foster relative to the optimal balance for the particular
task characteristics.

Strongly linked to the idea of appropriate and inappropriate goals is the idea of
insight problem solving. Insight problems are defined as those problems that require
restructuring of the initial problem representation to reach a solution (Chronicle,
MacGregor, & Ormerod, 2004). This is to say that participants, when placed in the task,
exhibit a general and (often unanimous) default representation of the task that is sub—
optimal. Successfiil performance on an insight problem requires the individual to break
away from the individual representation and search for a more functional representation.
This required shift in representation is what allows Miller, Lehman, and Koedinger
(1999) to make claims about the appropriateness or inappropriateness of a goal. In their
study those in the no goal and appropriate goal conditions were able to represent the task

differently to their benefit than those in the standard goal condition. Further, those in the

13

no goal condition were able to fully represent the task instead of just the areas of it
relative to specific outcomes.

Also in the field of problem representation, Vollmeyer, Burns, and Holyoak
(1996) showed that having a goal encouraged participants to deduce ways to simply meet
that goal without creating hypotheses about the underlying structure of the problem. Put
another way, those in the goal condition were encouraged by the goals to use a goal-
oriented or means-end method of solving the task (i.e. exploitation). Those in the
nonspecific goal condition had spent their time producing and testing hypotheses and
learning the structure of the task (i.e. exploration) such that when given a different goal
they were able to apply the same rules and reach the goal almost trivially.

Exploration/Exploitation and Goal-Setting Effects

This balance of exploration and exploitation is also present in some goal setting
research, although not specifically treated as such. In a study by Seijts and Latham
(2001), it was found that “a specific, difficult learning goal leads to higher performance
than urging people to ‘do their best’ on a task where people initially lack the requisite
knowledge to perform it” (p22). Conversely, it was found that “specific, difficult
outcome goals. . .appear to have a detrimental effect on performance during the learning
process” (p22). The finding by Seijts & Latham (2001, 2005) is consistent with the idea
that specific, difficult (performance) goals will decrease exploration on a task, as learning
is a form of exploration involving the collection of some new knowledge. This second
result -— that specific, difficult goals are detrimental during the learning process — is
consistent with Kanfer and Ackerrnan’s (1989) findings on the air traffic controller task.

Like the air traffic controller task, as well as a number of studies along this topic

14

(Kozlowski & Bell, 2006; Seijts & Latham, 2005), the task in Seijts & Latham (2001) is a
complex task where learning is required. Thus, while this result of specific, difficult
performance goals affecting performance dependant on stages of learning has been
shown on complex tasks where learning is required, little, if anything has been said of
predicted effects on less complex tasks or tasks where no learning is required. A more
basic explanation involving the promotion or inhibition of exploration by goals works not
only in these cases of complex tasks demonstrated by Seijts and Latham (2001), but also
in situations of simple tasks where exploration on task. — not just learning — can be
required or detrimental.

Moreover, the findings of Seijts and Latham (2001) require a shift to a
qualitatively different paradigm than what has been used in the traditional goal setting
studies that they are using as foundation. Mastery goals are qualitatively different in
structure and purpose than traditional performance goals. It was shown that mastery
goals can promote learning and exploration. In situations where learning and exploration
is useful mastery goals should be useful, and in situations where learning is not required
and/or exploration is detrimental mastery goals should not be useful. A discussion much
further into the realm of mastery goals is beyond the scope of this paper and the issue at
hand. The inconsistent goal setting effects found in traditional goal setting literature
(Burns & Vollmeyer, 2002; DeShon & Alexander, 1996; Early et a1, 1989; Huber, 1985;
Kanfer & Ackerman, 1989; Sweller & Levine, 1982) all have to do with performance
goals, and thus performance goals and their relation to ‘do your best’ goals is what must
be first explained. While Seijts and Latham’s (2001) work provides an explanation of

effects in complex tasks using mastery goals, the promotion and inhibition of exploration

15

by specific, difficult performance goals leads to the same predictions in traditional goal
setting theory from a more basic yet more encompassing standpoint, explaining not only
complex tasks but also simple tasks and all those in between. Moreover, it is done
without needing — but also allowing for — mastery goals.

Earlier in the paper, the results of Kanfer and Ackerman (1989) were shown to be
in direct contradiction to the findings of Wood, Mento, and Locke (1987) in that the goal
setting effect was moderated with no change in complexity. How, then, does exploration
and exploitation fare in explaining this result? Early in the task exploration is a useful
strategy as participants learn the rules of the task. Thus, if goals reduce exploration, they
will deleteriously affect performance. This is consistent with results from Kanfer and
Ackerman’s (1989) study one. Later in the task, afier all participants have explored
enough to learn the rules of the task, further exploration is not useful and exploitation is
the favored strategy. Thus, if goals reduce exploration, they will positively affect
performance. This is consistent with the results of Kanfer and Ackerman’s (1989) study
two.

Other tasks in the goal setting literature also fit into this framework. Sweller and
Levine (1982) used an experimental design in which the manipulation consisted of giving
the participant the location of the exit of a maze (southwest of the start of the maze).
This drove participants to exert all their effort at moving southwest while the path to the
exit actually existed to the northeast. Participants without knowledge of where the goal
was located had no choice to explore the maze, and outperformed the goal condition by

an overwhelming margin. So powerful was this effect, Sweller and Levine (1982) also

16

provided a replication, of which they noted: “The presence of a goal has again rendered
the problem essentially insoluble” (p 469).

In a study by Earley, Connolly, and Ekegren (1989), it was found that specific
difficult goals were harmful to initial performance on a complex task, similar to the
findings of Kanfer and Ackerman (1989). Earley, Connolly, and Lee (1989) completed a
study in order to both replicate these results and try to go one step further. The task in
this study consisted of a stock market prediction task in which participants were to
predict stock prices for 75 fictional companies based on performance numbers from
different departments. This is an example of a multiple cue probability learning task, or
MCPL task. Participants were placed in a specific, difficult goal condition or a ‘do your
best’ goal condition. Participants were also placed into one of three groups in terms of
strategy interventions. The first group received no strategy training, the second received
a list of ten possible prediction strategies (search restriction condition), and the third
received a step by step guide to developing their own accurate prediction strategy (search
training condition). It was found that, relative to others, subjects with specific difficult
goals did significantly worse in predicting stock prices on the first block of 15 trials
(exploitation before adequate knowledge has been acquired), but significantly better on
all trials following. It was also found that the strategy interventions had a significant
impact on performance and search for those in the goal condition, but not in the do your
best condition. This is perhaps due to the fact that those in the do your best condition
were already engaged in exploration of different strategy, while those in the specific,

difficult goal condition were not.

17

In another MCPL task, DeShon and Alexander (1996) implemented a series of
manipulations, the first between goal and do your best conditions, the second between
implicit and explicit strategy development. In this situation implicit pattern recognition
and learning can be conceptualized as an exploratory behavior, while a push toward
explicit learning can also be seen as a push toward exploitation. The outcomes of this
study suggested that when a task is processed explicitly, goals can be used to positive
effect by focusing attention on the task (exploitation). When the task is one where an
implicit strategy is best (exploration), goals will hurt performance.

Exploration and Reinforcement Learning

The distinction between exploration and exploitation is also illustrated in the field
of reinforcement learning. Reinforcement learning is “the problem faced by an agent that
learns behavior through trial-and-error interactions with a dynamic environment”
(Kaelbling, Littman, & Moore, 1996, p237). Task-participant interaction in
reinforcement learning can generally be described a situation in which a participant
receives a perceptual indication of the current state of some system, takes an action on
that system, and is presented with feedback as to what result the action produced. Over
time, the participant will choose actions which maximize the value of the result of
interest. For example, to use a task from the goal setting literature discussed earlier,
Kanfer and Ackerman’s (1989) air traffic controller task can be described in this way,
incorporating the balance of exploration and exploitation. At any given time the
participant can visually determine the state of the system of planes and runways, take

actions - such as landing a plane, and then see the result of that action on their score. The

18

participant may explore the rule set and learn about the task or apply whatever knowledge
is held at the moment to immediate performance.

While even tasks as complex as these can be described in this way, much of
reinforcement learning focuses on simpler tasks so that distinctions such as those between
exploration and exploitation can be made clear, and so that computer algorithms can be
written to solve them. One such simple task is the k-armed bandit problem (Berry &
Fristedt, 1985). This task is summarized by Kaebling et al (1996):

“The agent is in a room with a collection of k gambling machines (each called a

‘one-armed bandit’ in colloquial English). The agent is permitted a fixed number

of pulls, h. Any arm may be pulled on each turn. The machines do not require a

deposit to play; the only cost is wasting a pull playing a suboptimal machine.
When arm i is pulled, machine i pays off 1 or 0, according to some underlying

probability parameter, pi, where payoffs are independent events and the pis are
unknown.” (p234)

In a situation where k is greater than one, the agent is faced with a dilemma. If
the agent believes that one of the machines has the best underlying pay probability,
optimal strategy should dictate that all pulls be on this machine (Kaebling et al, 1996).
However, there is a chance that a different machine has a better pay probability and that it
is simply the case that not enough information has been collected to know it. Thus, the
agent can choose to explore the other machines in order to collect new information at the
cost of acting on what it believes to be optimal. This is the struggle between exploration
and exploitation.

Another fairly simple task arises out of foraging theory (Pyke, 1984), mentioned
earlier. Foraging theory examines the behavior of agents (often animals) who must
interact with their world in order to find and utilize resources for survival. However,

foraging theory is also applicable to human behavior, both in our distant past as hunter

l9

gatherers (Bird, Smith, & Bird, 2001) and even in recent times in terms of decision
making (Rode, Cosmides, Hell, & Tooby, 1999).

In foraging theory, an agent must exploit resources in order to survive, but also at
the cost of depleting those resources. As such, the agent must also explore the world in
order to gain knowledge about previously unknown resources that may be more or less
rich or abundant than those already known. Thus, the agent is faced with a dilemma: the
choice of exploiting known resources versus the choice of exploring for new resources.

What determines how this balance will be decided? In a situation without
constraint it is easy to imagine that a person will try to reach a balance that allows for
investment in both exploration of alternatives and exploitation of presumed optimality.
However, constraints on the agent may cause the balance to swing one way or the other.
As stated earlier, it is suggested that specific, difficult goals are one such constraint.

Summary and Hypotheses

How then, do goals produce these constraining effects? Recall that goals “direct
attention and effort toward goal-relevant activities and away from goal-irrelevant
activities” (Locke & Latham, 2002, p706). Exploration is cases “for which there appears
to be no goal object or condition to and for which the organism responds” (Fowler, 1965,
p23). Exploratory behavior is goal-irrelevant behavior, such as wandering away from
resources in a foraging theory world. Goals should inhibit this behavior and instead
swing the balance of exploration and exploitation to the side of exploitation, as shown in
numerous studies in the previous section (DeShon & Alexander, 1996; Barley, Connolly,

& Ekegren, 1989; Barley, Connolly, & Lee, 1989; Kanfer & Ackerman, 1989; Miller,

20

Lehman, & Koedinger, 1999; Seijts, Latham, Tasa, & Latham, 2004; Sweller & Levine,
1982; Vollmeyer, Burns, & Holyoak, 1996). In short, goals inhibit exploration on a task.

HI : The presence of a diflicult, specific goal, relative to a condition of a ‘do your
best ’ goal, will inhibit exploratory search behaviors, negatively impacting the pattern
and/0r quantity of such behavior.

Further, this effect on exploratory behaviors will manifest itself as a goal setting
effect. As it is found in the literature, this goal setting effect will sometimes be strongly
negative, sometimes be strongly positive, and sometimes be differing levels in between.
The direction of the effect (positive, negative, or neutral) will depend on characteristics
of the task, and can be well conceptualized in the framework of insight problems as
discussed above (e. g. Chronicle, MacGregor, &Ormerod, 2004):

H2a: If the default method for completing the task is the most effective, goals will
limit exploration and direct effort toward this method, creating a positive goal setting
effect.

Conversely:

H2b: If the default method for completing the task is not the most eflective and
exploration is required to find the most eflective method, goals will create a negative
goal setting eflect by limiting exploration and thus lessening the chances of finding the
most effective method.

Individual Difference Effects

There are individual difference constructs in psychology that are linked to the

ideas that have been presented so far as the foundation of this basic process. The

constructs that are seemingly most related to these ideas are those constructs of curiosity

21

and goal orientation. There are reasons to believe that these individual differences in
participants may cause effects in concert with the above hypotheses. These proposed
individual difference constructs will be discussed, and predictions involving their effects
will be hypothesized.
Curiosity

Curiosity has been defined as “a desire to acquire new knowledge and new
sensory experience that motivates behavior” (Litrnan & Spielberger, 2003, p118). It has
also been recognized as a critical motive that influences human behavior (Loewenstein,
1994), and has been further conceptualized as “information seeking, or cognitive
curiousity, which stimulates information-seeking, exploratory behavior” (Reio, Petrosko,
Wiswell, & Thongsukmag, 2006). This linkage to exploratory behavior is of use to this
study through the proposed linkage between exploration and goal setting. As stated, the
study of exploratory behaviors has long been intertwined with the study of curiousity as a
possible drive of these behaviors (Fowler, 1965). Curiosity has also been shown to be a
stable individual difference trait that can be measured (Kashdan, Rose, & Fincham,
2004). In all, some people may be generally more curious in their behaviors than others,
and thus may naturally engage in more exploratory behaviors (Litrnan, Hutchins, &
Russon, 2005; Reio, Petrosko, Wiswell, & Thongsukmag, 2006). Thus:

H3: Those who score higher on ratings of trait level curiosity will engage on
average in more exploratory behaviors than those who score lower. If this is found, trait

level curiosity will be controlled for in above eflects involving exploratory behavior.

22

Goal Orientation

Goal orientation is generally defined as the way that individuals interpret and
approach situations in which they are expected to reach an achievement outcome
(Dweck, 1986). Early conceptualizations of the topic outlined two distinct styles —
learning and performance goals. Dweck (1986) defined learning orientation as that of
individuals that approach a task with the intent of learning and increasing competence for
their own sake, and performance orientation as that of individuals who approach a task
with the intent of gaining positive (or avoiding negative) judgments of their result.
Performance goal orientation has fiirther been distinguished between the ideas of gaining
positive or avoiding negative judgments of performance such that performance approach
goal orientation has been defined as “the desire to prove one’s competence and to gain
favorable judgments about it” (V andeWalle, 1997, p. 1000), and performance avoid goal
orientation has been defined as “the desire to avoid the disproving of one’s competence
and to avoid negative judgments about it” (VandeWalle, 1997, p. 1000).

For the purposes of this study, learning goal orientation and performance avoid
goal orientation are of note. Learning goal orientation is generally behavior that is not
motivated toward some end state — simply behavior for the sake of behavior. It is
important to note the distinction between behaviors driven by an individual’s learning
goal orientation and those driven by learning goals (Seijts & Latham, 2001). This links
well to the ideas of exploratory behaviors in that they are behaviors without motivation
stemming from a concrete end state. Performance avoid goal orientation is also
potentially related through its effects on fear of failure (VandeWalle, 1997). Individuals

with goals have concrete methods by which they can evaluate themselves (against the

23

goal) while those in do your best conditions do not (accept against their own prior
performance).

Due to the non—outcome focused nature of exploratory behavior - the idea of
exploration simply for the sake of gathering new knowledge, it is possible that there may
be a linkage between exploratory behavior and goal orientation. Hints of this linkage can
be found in the goal setting literature (Seijts et a1, 2004). Specifically:

H4a: Those high on learning orientation will engage in more exploratory
behaviors than those low on learning orientation.

H4b.‘ Those high on performance avoid orientation will show a stronger
inhibiting effect of goals on exploratory behavior due to desire to reach those goals and

avoid failure.

24

METHOD AND RESULTS
Study 1: Pilot

To explore the feasibility of the proposed manipulations and hypotheses a small
pilot was completed using a reduced version of the final method on a different but
conceptually similar task.
Participants

Participants consisted of six graduate student volunteers from the psychology
department of a large Midwestern university. None of the participants had knowledge of
the task or principles underlying the task prior to the experiment. Participants were
exclusively Caucasian and predominately male (66%).
Procedure

The task in this study was a k-armed bandit with four arms presented to
participants through a computerized interface. Upon arrival at the lab, participants
received a brief training on the task. This training involved explanation of how to operate
the simulation as well as background information informing them that they should
imagine themselves as a member of the human resources division of a large company.
As a member of human resources, it was their job to hire workers from four different
temp agencies — each of these agencies was represented in the task as an arm of a k-
armed bandit as summarized by Kaebling et a1 (1996) and discussed above. In this
situation the four arms of the bandit are represented by four buttons. In order to remain
consistent in terminology these four buttons will be referred to as the ‘arms’ of the task —

as that is what they are representing in a reinforcement learning framework. A view of

25

the screen presented to participants in the goal condition and clearly showing these four
buttons can be seen in figure 4.

Put simply, each participant was faced with four arms of which at any given point
one could be activated. This activation led to success (score increments by one) or failure
(score stays the same) based on predetermined underlying probabilities of the task. Over
the course of the experiment participants completed 20 trials of the task, each task
consisting of 50 ‘pulls’ on the k-armed bandit. As stated, the k-armed bandit in this
experiment had four arms (hereafter referred to as arm A, arm B, etc.). The probabilities
on arms A, B, and D were fixed throughout the experiment at 70%, 20%, and 40%
probability of return, respectively. The probability of arm C was differential in such a
way that it was a poor arm in the first half of the experiment (30% return for trials 1-10),
but the best arm in the second half of the experiment (90% return for trials 11-20).

Participants were also placed in one of two conditions for the experimental
manipulation. In both conditions, text was displayed above the four bandits throughout
the entire task. This text is the method by which the manipulation was enacted. In the
first condition (do your best condition), the text read ‘Do your best to maximize score.’
In the second condition (specific, difficult goal), the text read ‘Achieve a score of 33 out
of 50.’ A score of 33 out of 50 (i.e. a 66% return rate) represents a specific, difficult goal
in the fact that in order to achieve it (in the first 10 trials), one must invest nearly all their
pulls on the highest probability arm (arm A). No other arm in the first ten trials would
deliver on this goal. As such, the first half of the task rewards exploitation on arm A. In

the second half of the experiment, exploitation on arm A is not an optimal strategy, as

26

exploration can reveal that arm C is in fact a better choice and where effort should be
invested.

After the task participants were also verbally debriefed by the experimenter as
well as posed a series of task knowledge questions such as ‘what arm had the best
probability?’

Results and Discussion

Two variables of interest were examined as part of this pilot study: probability of
pulling any given arm on any given trial, and performance. Probability was computed by
counting the number of times a participant pulled any given arm during a trial, and then
dividing by 50. This yielded four percentages for each trial. Performance was the
number of returns a participant accumulated during a trial.

Results for probability were consistent with hypotheses, as a repeated measures
analysis showed that the three way interaction of time, condition (goals vs do your best),
and probability of arm choice was significant (F(57,304) = 2.359, p < 0.001). This
finding is also apparent in figures 1 and 2. Figure 1 displays the probabilities for the ‘do
your best’ condition and reveals that participants in this condition were more likely to
find the better arm in the second half (compared to the goal condition represented in
figure 2).

As well, results for performance were also consistent with hypotheses. While a
repeated measures analysis showed that the interaction of performance and condition was
not significant (F(l9,76) = 1.25, p = 0.239; represented in figure 3), significant
differences were found on the last two trials after the participants in the ‘do your best’

condition explored enough to determine the better arm to pull (F(1,4) = 11.63, p < 0.05, F

27

= 10.321, p < 0.05). Additionally, while the first few trials were marked by exploration
on the task in both conditions due to the need for all participants to learn the probabilities,
the last few trials of the first half represented a situation where exploitation was the
favored strategy. Performance differences in these trials reached marginal significance
on one of the trials (trial 8, F(1,4) = 4.985, p = 0.089), suggesting that a difference might
be present and require more statistical power to reveal (or dismiss).

The results of verbal debriefings also revealed that of the participants in the ‘do
your best condition’, all were able to respond to the question of ‘which was the best arm’
with the split answer of A in the first half, C in the second. Only one participant in the
specific, difficult goals condition was able to make close to this distinction, stating that A
seemed better at first, but that at some point it seemed like C might have gotten to about
the same probability as A. The other participants in the specific, difficult goals condition
believed that A was the best arm throughout the entire task.

Overall, this pilot study supports the idea that specific difficult goals limited
exploration on this k-armed bandit task; these initial results are strongly consistent with
the proposed hypotheses. However, the bandit task as presented is incapable of
addressing the concern of the need for a complexity manipulation that produces results
counter to that of Wood et a1 (1987). In order to address the concerns, it is necessary for
the task to be manipulated in both complexity and need for exploration in a way that does
not confound the two. As well, goals must have a way to work on the task through their
effects on effort. The bandit task, since there are a set number of pulls in each trial, does

not allow differential effort to manifest itself over different conditions. Thus, we have

28

determined that a similar task from the area of exploration/exploitation literature would
be better suited to find the results we are proposing.
Study 2: Pilot 2

The task chosen to better examine the relationship between exploration and task
complexity is a computer simulation modeled after the ideas of foraging theory. In the
simulation, the participant interacts with a two-dimensional world which they can
navigate in search for resources (in foraging theory these would be food, water, etc). The
world, except for a small area within a radius around the participant (in this case 3
squares in any direction), is hidden from view until explored. The participant is able to
move in any direction with the arrow keys to explore the world and find resources, which
are scattered throughout the map in patterns dependent upon experimental manipulation.
Once a participant encounters a resource they then must devote time to collecting it
(using the spacebar), accumulating a score. There is both a move delay and a collect
delay such that participants cannot move again afier moving for .5 seconds, or collect
again after collecting for .2 seconds. The particulars of the experiment are described
below.
Participants

Data was collected from 47 undergraduate students at a large Midwestern
university. Participants were roughly evenly split on gender and predominately
Caucasian, all between the ages of 18 and 24. Participants received credit in the

psychology research pool for their participation.

29

Design

The purpose of this pilot was to determine at what level the goals should be set for
conditions involving difficult specific goals, as well as show that the different world
maps being used would differ correctly not only on objective complexity, but also on
perceptual complexity, while not differing on difficulty.

For these purposes, all participants engaged in a ‘do your best’ version of the task.
Additionally, participants were placed into one of three between subject conditions, each
consisting of a different version of the world with which they interacted. Each world
consisted of a 20x30 square space; the differences arose in the placement of resources
throughout each individual world.

In condition 1 (World 1, figure 4), there were small pools of resources within
sight of each other near the place where the participant starts. The participant starts near
the ‘bottom’ of the world, at the center. In this way they are surrounded by the dispersed
pools of resources. This section of small pools of resources accounts for slightly more
than half of the task’s world. In the rest of the world there is one large (relative to the
smaller pools) pool of resources that the participant can work on without moving. This
pool can be said to thus be a richer source of resources.

In condition 2 (World 2, figure 5), the world was mirrored such that participants
still start at the ‘bottom’ of the screen, but this part of the world is now empty accept for
the pool of rich resources located to their immediate right. They are one square away
from being in sight of the larger pool of resources, and the small pools of resources are on

the top half of the screen.

30

In condition 3 (World 3, figure 6), a complexity manipulation was used relative to
the first condition. Participants started in the same place as in the first world, at the
center of the ‘bottom’ of the screen, surrounded by dispersed resources. In order to make
the task more complex in the most clear and powerful way possible, dynamic complexity
was manipulated. It was manipulated by making the pools of resources non-stationary
across trials. The smaller pools of resources change position within their half of the
screen, as does the larger pool of resources on the top half of the screen. This occurs by
random placement with each of three areas of the screen being 33.3% likely of having the
pool of resources on any given trial. As the resources on the bottom half of the screen
were generated pseudo-randomly, a new set was simply generated for each trial.

As the participant moves around each of these three worlds more and more of the
world becomes visible. Any square that has been seen during that trial remains visible
for the remainder of that trial, but participants return to the start and to their initial scope
of vision at the start of each trial. Examples of the amount of the world that can be seen
early and mid trial by participants can be found in figures 7 and 8.

Procedure

Upon arrival at the lab, participants were randomly assigned to one of the three
conditions of the experimental design. Participants then received a brief training on the
task (approximately 5 minutes). This training informed them of what they would be
doing during the session (how many trials they would complete, etc), and how to move
around the world and collect resources. This training was uniform across conditions.
Participants then engaged in a one-minute practice trial on a much smaller world in order

to be sure that they knew how to move and collect resources properly. Following this,

31

participants completed five two-minute trials on the task as outlined above. At the end of
each trial, participants received score feedback on the previous trial.

Upon completion of the experiment, participants completed a series of exit
measures including a number of measures assessing both the complexity and difficulty of
the different versions of the task.

Measures

Perceptu_al Complexity — Perceptual complexity was measured using a three item
scale containing one question about each of the different facets of complexity as outlined
by Wood (1986). These three items, measuring component, coordinative, and dynamic
complexity, respectively, are: “Rate the number of unique physical actions that had to be
completed in order to accomplish the computer task you completed during this
experiment”, “Rate the degree to which actions which you completed during the
computer task were linked to each other in terms of frequency, timing, and coordination”,
and “Rate the degree to which characteristics of the task and the actions best suited for
the task varied from trial to trial.” Each item was measured on a 5-point scale with
tailored anchors.

Percepmflifficultv — Perceptual difficulty was measured using a one item scale
meant to address how difficult and challenging participants believed the task was. This
item was: “How difficult was the computer task in this study?”, with anchors on a 5-point
scale ranging from very low difficulty to very high difficulty.

Results
The purpose of this pilot was twofold. First, participants were placed in this

version of the task in order to set a goal for those in the specific, difficult goal condition.

32

Keeping with the goal setting literature, a goal was calculated for the 85th percentile of
those in the do your best condition. However, in order to limit the potential effects of
score boost by exploration as would occur later in the task, the goal was calculated only
on the first trial of the do your best condition. For world 1 this produced a goal of 107.9,
which for practicality was rounded to 108. As there are 28 patches of resources, each
containing 4 resources on the bottom of the world — that which is easily exploitable
without exploration — this goal of 108 out of a maximum of these 112 is not unreasonable
or impossible, but still quite difficult. For world 2 this produced a goal of 142.6 (rounded
to 143), which is reasonable due to the fact that participants start on the part of the world
with a much richer area of resources. For world 3 this produced a goal of 108.0, which is
in practicality identical to the goal set for world 1. This is likely due to the fact that
without exploration world 1 and world 3 are both randomly generated to identical
specifications in the first trial.

The second purpose of this pilot was to show that results of perceptual complexity
would be consistent with our assessments of objective complexity. It was proposed that
neither world 1 and world 2 should be more or less complex than the other — the
transformation between them is a non-complexity based transformation. However, world
3 was transformed from world 1 with a complexity based transformation, specifically a
transformation involving dynamic complexity. Thus, world 1 and 2 should be
perceptually similar in terms of complexity, while world 3 should differ from both in
dynamic complexity. This is precisely what was found.

A one-way ANOVA on all three groups showed no differences on component

(F(2,56) = .464, ns) or coordinate (F(2,56) = .095, ns) complexity, but did show a

33

difference on dynamic complexity (F (2,56) = .3.828, p < .05). Post hoc tests revealed
that this effect was manifested in mean differences between world 1 and 3 and world 2
and 3, with participants in world 1(M= 2.38, SD = .94) and world 2 (M: 2.31, SD = .79)
reporting significantly lower dynamic complexity than those in world 3 (M = 3.13, SD =
1.06).

Additionally, the result that perceptions of difficulty did not differ between the
three groups (F(2,56) = 1.474, ns) was found. Thus, while the sample size is relatively
small, it hints that our complexity manipulation is not confounded with a difficulty
manipulation.

Study 3: T est of Goal-Setting Eflects
Participants

Data was collected from 332 undergraduate students at a large midwestem
university. Of these participants, 47 were those of the previous pilot study included
because of the identical nature of their cells of the experimental design. Participants
received credit in the psychology research pool for their participation or extra credit in a
psychology class. Of these 332 participants, 20 participants were removed from the
dataset for having large amounts of incomplete data on the task. This was due to a
number of possible random factors such as computer error. Of the remaining 312
participants, 2 were removed for having a score of less than 50 on any trial. As only two
participants received scores this low on any trial it is to be taken as a showing of low
motivation or a misunderstanding on the task, not of any other important process. Thus,

the final data set consisted of 310 participants.

34

 

WE

STU

WC

PM

on]

0111

W0:

Ag;

Eat

Of these 310 participants, the majority (65%) were female. As well, the majority
were Caucasian (81%). The mean age for participants was 20.05 years, and the majority
was between the ages of 18 and 22.

Design

The design of the study was a three by two between subjects. The first between
subjects experimental manipulation was identical to that of the pilot studies — one group
received difficult specific goals (difficult specific goal condition) throughout the
experiment while another received do your best goals (do your best goals condition).
These goals will be displayed on the screen in the same location as the text ‘do your best’
was displayed to those in the ‘do your best’ goal condition of the initial pilot. The second
between subjects manipulation was identical to the manipulation in the second pilot
study. That is, participants performed the task in one of three worlds, world 1 (figure 4),
world 2 (figure 5), or world 3 (figure 6).

Procedure
Trait level measures (e. g. trait level curiosity, goal orientation) were measured
online prior to the study in order to obtain results unbiased by any lab experience. This
online measurement took place immediately after participants signed up for the study,
and was thus separated from the lab session by anywhere from one day to one week’s
worth of time.
Upon arrival at the lab, participants received a brief training on the task identical
to that of the second pilot study, as well as an identical one-minute practice session.
Again identical to the pilot participants next completed five two-minute trials on the task.

Each of these trials was followed by score feedback and a series of process measures. A

35

final set of measures (perceptual complexity, perceptual difficulty) was completed after
the last set of process questions.

In conditions 1 & 2 (world 1, figure 4), there were small pools of resources within
sight of each other near the place where the participant starts. By devoting time to
moving from each of these to the next, the participant stood a chance of reaching the
difficult goal presented to them. However, as found in the pilot, this was still a difficult
prospect. This section of small pools of resources accounted for roughly half of the
task’s world. On the other half of the world, there was one large (relative to the smaller
pools) pool of resources that the participant could work on without moving. Thus, in this
first version of the task, those that explored and found the larger pool should outperform
those that spend their time exploiting. Thus, if difficult specific goals do in fact limit
exploration as hypothesized, this should produce a negative goal setting effect, as those
with goals will exploit and those without (do your best conditions) will explore
(hypothesis 2b).

In the second version of the task (conditions 3 & 4, world 2, figure 5), the world
was mirrored such that participants started near the larger pool of resources. Thus, in this
version of the task - with equal complexity to the first - any exploration (time spent not
collecting resources) should lead to a decreased score. Those that exploited on the task
(difficult goal condition) should thus outperform those that spent time exploring. Again,
if difficult specific goals limit exploration then those that are given goals will exploit and
those without will explore. This should produce a positive goal setting effect (hypothesis

2a).

36

For the third version of the task (conditions 5 & 6, world 3, figure 6) a complexity
manipulation was used relative to the first condition. In order to make the task more
complex in the most clear and powerful way possible, dynamic complexity was
manipulated. It was manipulated by making the pools of resources non-stationary across
trials. The smaller pools of resources changed position within their half of the screen (by
trial), as did the larger pool of resources on the top half of the screen. This occurred by
random placement, with each of three areas of the screen being 33.3% likely of having
the pool of resources on any given trial (figure 6). Thus, while participants gain some
benefit of exploring and finding the pool of resources, they can also never gain concrete
knowledge about its position through this exploration. Through this effect the benefits of
finding the pool (conditions 1 & 2) are offset by the price of finding it (i.e. time spent
exploring). While this has little effect on those that exploit (some small pool should by
chance always be in the participant’s sight when they start), it will have an effect on those
that explore in that more time is required (on average) while exploring to find the larger
pool. This should lower the score of those that explore, thus shifting the task to one that
again favors exploiting. Since it is hypothesized that those that are given specific,
difficult goals will be those more likely to exploit, this should shift the benefits of the
task back toward their favor — on a task that is unquestionably more complex by Wood
1986 criteria than the simpler version that produced a negative goal setting effect. It is
unclear then if this will be powerful enough to produce a positive goal setting effect or if

it will simply shift the negative goal setting effect to lesser negative or neutral.

37

Measures

T_rait level gtfil orientam — Goal orientation is defined as the method and
motivation behind individuals’ work toward different types of goals. Two aspects of goal
orientation are proposed to have effects in this study; learning goal orientation and
performance avoid goal orientation. Both were measured using a nine item measure
developed by VandeWalle (1997). Five items are designed to measure learning goal
orientation and four items are designed to measure performance avoid goal orientation.
Example items are “I often look for opportunities to develop new skills and knowledge”
(learning) and “Avoiding a show of low ability is more important to me than learning a
new skill” (performance avoid). This measure has been shown to have high internal
consistency (a=.88 for learning, or=.83 for performance avoid), and reasonable test-retest
reliability (a=.66 for learning, a=.57 for performance avoid). Validity of the measure has
also been shown through correlations with similar measures (VanderWalle, 1997).

Trait level curiosity — Curiosity was measured using a four item scale from
Kashdan, Rose, and F incham (2004) designed to measure a trait level model of curiosity
involving exploration. Example items include “Everywhere I go, I, am out looking for
new firings or experiences” and “I frequently find myself looking for new opportunities to
grow as a person (e. g. information, people, resources)” (Kashdan, Rose, and Fincham,
2004). This measure has been shown to have a reliability of or=.74, and a test-retest
reliability over a month period of r=.78, p<.001. It has also been shown to possess
convergent validity with other curiosity measures, openness to experience, and positive
affect and discriminant validity with conscientiousness and agreeableness (Kashdan,

Rose, & Fincham, 2004).

38

sale
by \I
had
6pr
con
and
the

tai.

Perceptual Complexity — Perceptual complexity was measured using a three item
scale containing one question about each of the different facets of complexity as outlined
by Wood (1986). These three items are: “Rate the number of unique physical actions that
had to be completed in order to accomplish the computer task you completed during this
experiment”, “Rate the degree to which actions which you completed during the
computer task were linked to each other in terms of frequency, timing, and coordination”,
and “Rate the degree to which characteristics of the task and the actions best suited for
the task varied from trial to trial.” Each item was measured on a 5-point scale with
tailored anchors.

Perceptual Difficulty — Perceptual difficulty was measured using a three item
scale meant to address how difficult and challenging participants believed the task was.
These three items are: “How difficult was the computer task in this study?”, “How
challenging was the computer task in this study?”, and “How easy was the computer task
in this study?” (reverse-scored). Each item was measured on a 5-point scale.

Exploration (operationalization) —- Exploration can be operationalized in two
different ways: 1) how much of the map (by percentage) was visible to the participant,
cumulatively across trials, during any given trial, and 2) whether or not the participant
crossed the (conceptual) line between the side of the map containing the dispersed
resources and the side of the map containing the rich resources. From this point on these
variables will also be referred to as EXl (percentage of map visible) and EX2 (crossing).
Where appropriate and in situations when these two variables do not differ in their effects
they will be addressed simply as exploration. When they do differ, or if the effects are

similar but differ in interpretation, results will be discussed separately.

39

Results

Results will be ordered as follow. To begin, manipulation checks involving
complexity and benefit of exploration will be examined. Following this, individual
difference effects will be tested in order to determine whether or not these individual
differences should have a role in further analyses. Following this, full tests of the
proposed model will be tested.

Manipulation Check — In order to test whether or not participants found any world
to be more complex or difficult than the others, a series of analysis of variance was
completed on self report complexity and difficulty measures by world. While an effect
was found for dynamic complexity in pilot data such that participants self reported world
3 to be more dynamically complex than worlds 1 and 2, this effect was not found in the
full sample. Further, no effects of any level of perceptual complexity or perceptual
difficulty were found.

Additionally, in order to determine if exploration was more or less useful in any
world as predicted (exploration beneficial in worlds 1 and 3, detrimental in world 2), a
number of linear regressions spilt by world were performed predicting score at each trial
by the amount of exploration by that trial. The results of these regressions can be found
in table 1 and show that exploration is positively related to performance in worlds 1 and 3
and negatively related to performance in world 2, supporting the need for exploration
manipulation.

Main An_alvses - In order to test hypotheses 3, 4a, and 4b, a series of linear
regressions were performed. These analyses tested the predictive power of trait level

curiosity, trait level learning goal orientation, and trait level performance avoid goal

40

orientation on exploration. Both operationalizations of exploration were tested, with the
‘gone across’ variable in table 2 and the cumulative exploration variable in table 3, both
tested at each individual trial. No significant effects were found. Thus, hypothesis 3, 4a,
and 4b are not supported, and these individual differences will not be included in further
analyses.

In order to begin testing remaining hypotheses, repeated measures analysis was
first performed on performance (score) and exploration (as outlined above) by trial, goal
condition (do your best vs. specific difficult), and world (world 1, 2, and 3). For
performance there are a number of significant effects, including the three way interaction
of time by goals by world (F(8,1216)=3.01, p<.01). This three way interaction is not
significant for exploration (EXl, F(8,1216)=1.48, ns; EX2, F(8,1216)=1.06, ns). The
remaining results of these two analyses can be found in table 4 and will be examined in
turn as a means of explaining these more complex interactions and testing the predicted
hypotheses.

First, in terms of time effects, participants’ performance was generally increasing,
with a mean score of 106.19 on the first trial and a mean score of 135.05 on the last trial.
Conversely, participants’ exploration was generally increasing over time, as more
participants were more likely to explore the map as time went on. While performance is
explained by participants getting better on the task, exploration likely rests on the fact
that at least some participants will see something new each trial through exploration. As
participants see new parts of the map their cumulative exploration increases, as does the

chance that they will cross the dividing line between sections of the worlds.

41

There is also an effect of world on both performance and exploration. Further
examination of this effect shows that participants in world 2 scored significantly higher
on average than those in world 1 (F(1,2ll)=37.23, p<.01) and world 3 (F(l,209)=39.99,
p<.01). This effect is likely due to the fact that participants in world 2 start near the
richest area of resources. Thus, even if they do explore, they are more likely to return to
this area to maximize their score. Further, those collecting the rich resources simply have
less distance to travel to get to them. This score differential is also consistent with pilot
data on do your best only conditions (pilot study 2). Additionally, differential
exploration effects appear. Participants in world 2 are more likely to travel across to the
other side of the map than in conditions 1 or 3. This is due in part to the nature of the
worlds — the division between the distributed resources and the rich resources is at the
two-thirds point of the map. That is, the area with distributed resources takes up two-
thirds of the map, and the area with rich resources takes up one-third. Thus, participants
in world 2 start closer to the line due to the fact that they start on the side with rich
resources. In terms of cumulative exploration, participants in world 3 tend to explore
more than participants in the other worlds. This is likely due to the changing nature of
the world — participants need to explore more areas of the world in order to continue
doing the same thing, whereas participants in the first two worlds can maintain the same
pattern across trials.

The non-significant effect of goals on performance across all worlds is to be
expected — there are different predictions for each world in terms of goal-setting effects.
While in world 1 participants with do your best goals should perform better than those

with specific, difficult goals, participants with ‘do your best’ goals in world 2 should

42

perform worse than those with specific, difficult goals, effectively canceling out effects
across conditions.

In a direct test of hypothesis 1, the effect of goals on exploration was tested. This
effect is in the predicted direction; those given difficult specific goals explore less (as
measured by both of the exploration variables) than those given do your best goals.
While it is not measured directly, it can be assumed that those given specific, difficult
goals exploit more than those given ‘do your best’ goals. These results are supportive of
hypothesis 1.

The next result is the interaction of time by world on performance and
exploration. In terms of performance this result is somewhat unclear in interpretation. It
appears that world 2 has a slightly larger slope over time than world 1 or world 3, and
world 3 has less consistent slope over time. This first effect is due to the fact that people
who eventually find the rich resources can score higher in world 2 than in the others. The
second effect is likely due to the shifting of the rich resources, as there is a chance that a
participant might find them right away during a trial or spend opportunity cost again
searching for them. The effect on exploration seems driven at the ‘gone across’ variable
by earlier effects of world 2 (participants more likely to cross early in the experiment),
and at the cumulative exploration variable by world 3. Specifically, those participants in
world 1 and world 2 seem to slow or stop exploring new parts of the map by the third
trial, whereas participants in world 3 continue to keep exploring new areas. This is due to
the fact that those that have found the rich resources need to search new areas in order to
find it. They pay for this continuing need for exploration in continuing opportunity costs

of it.

43

The time by goals interaction on both performance and exploration is also
somewhat difficult to interpret. Those given ‘do your best’ goals seem to have a slightly
higher slope in terms of cumulative exploration and going across the map over time,
though in terms of performance these same participants seem to have a more varied slope
over time. It is suspected that without taking into account what is happening in each
different world this result may simply be nonsensical.

In a direct test of hypotheses 2a and 2b, analysis was performed to test the effect
of a world by goals interaction on performance. The effect on performance is significant
and shows that in world 1, participants perform better when given difficult specific goals
relative to do your best goals, whereas in world 2 participants perform worse when given
specific, difficult goals relative to ‘do your best’ goals. These two results are supportive
of hypotheses 2b and 2a, respectively. While no direct hypotheses have been made about
world 3, it would stand to reason from Wood, Mento, & Locke (1987) that due to the fact
that world 3 is simply a more complex version of world 1, the goal setting effect in world
3 would be predicted by their results to be tending toward lower positive or greater
negative than whatever the goal setting result is in world 1. In world 3 there is no
difference in score between those given ‘do your best’ and specific, difficult goals. This
is likely due to the fact that the effects of performance play out over time with the gap
between goal conditions widening with each trial. Since those participants in world 3
must effectively re-learn the world every trial they never have a chance for this effect to
play out — thus a neutral goal setting effect is the result. This is counter to the prediction

of Wood, Mento, & Locke (1987) — that an increase of complexity would shift a goal

44

setting effect in the negative direction — and consistent with exploration as an explanation
of goal setting effects.

The effect of world by goals on cumulative exploration is not significant, but this
is to be expected. There is no reason that the effects of goals should differ by world. The
significant effect on exploration (in terms of the ‘gone across’ variable) seems to again be
driven by those participants in world 2. Those in the goal condition who spend their time
exploiting are somewhat likely to exhaust the rich resources before the very end of the
trial, at which point they will cross to the side with distributed resources. If this is in fact
the case, it would represent continuing exploitation of the world, and not a shift to
exploration, as they are only exploring as a means to continue what they know how to do.

One final exploratory analysis was completed on a subset of the participants of
world 3. It was observed during the coding of data that some participants who had found
the rich resources on previous trials crossed on following trial to where they once were in
order to collect them. Finding the rich resources no longer in the spot they last left them,
participants had a choice to explore more of the top. This increased exploration was at an
opportunity cost of not collecting for an additional and unknown time period, if anything
was even ever to be found. Participants may have only found the rich resources the trial
prior, or thought that removal of them was a valid possibility of the task design. Put
simply, participants were placed again in a situation where they were unsure if
exploration would pay off. Participants in this world knew that they could return to the
bottom and exploit the dispersed resources there, as they would have had to have passed
through them from the start to the top of the screen. As a further test of our hypotheses it

should be the case that this new subset of participants should be subject to the same

45

effects — those who are given difficult specific goals should be more inclined to return to
exploiting and those given do your best goals should be more inclined to continue
exploring. Statistical power is diminished by the fact that this analysis is on groups
within subsets of the population on only one of the worlds (sample size for this analysis
in trial 1 is only N=8), but as the trials progress more and more participants are likely to
be crossing sides of the map.

The only marginal effect found is on trial 5 (F(1,68) = 2.96, p=.09), but it is in the
right direction — 47% of those in the difficult specific goal condition returned to the
dispersed resources compared to only 27% in the do your best goal condition. In effect,
the difference between exploring and exploiting in the difficult specific goal condition
was akin to a coin flip, as each participant had a roughly half and half shot of continuing
to explore. Chances of returning in the do your best goal condition were cut almost in
half, as roughly 3 out of every 4 participants chose to stay and explore. Further, this
impacted these participants scores in a way which produced a negative goal effect (as
should be the case where exploration is beneficial); those who chose to continue
exploring outscored those who returned to the dispersed resources by a mean difference
of over 50 points (F(1,68) = 48.10, p<.01). While this is not a definitive replication of
the above results, it suggests that with these results can be generalized to other situations
where individuals face choices between exploration and exploitation.

Discussion

In summary, a number of hypotheses were supported and a number were

unsupported. The hypothesis relating difficult specific goals to a decrease in exploration

were supported - those participants given difficult specific goals explored less than those

46

given do your best goals. Further, by changing the nature of the world in order to alter
the task’s need for exploration differential goal setting effects were created. In the first
world, where exploration was beneficial, a negative goal setting effect was found. This
effect was reversed in the second world where exploration was detrimental, producing a
positive goal setting effect, and negated in the third world where exploration was started
anew every trial, producing no goal setting effect. However, those hypotheses relating to
individual differences as predictors of exploration were unsupported. Trait level
curiosity, trait level learning goal orientation, and trait level performance avoid goal
orientation failed to predict exploration at any trial of the experiment.

Globally, these results speak directly to the purpose of this paper, namely to
“show that the way in which specific difficult goals may be causing both positive and
negative goal setting effects is through their effects on exploratory behavior”.
Complexity was not a manipulation between world 1 and world 2, but rather by
manipulating the need for exploration in those worlds was the goal setting effect reversed
from a negative to a positive effect. When a complexity manipulation was put in place —
making the third world more complex relative to the first world — the negative goal
setting effect that had been found was neutralized, contradicting the current stance on
goal effects. From Wood, Mento, and Locke (1987) the goal setting effect should move
fi'om high positive to low positive when complexity is increased, not from negative to
neutral.

In short, and in summary, difficult specific goals were shown to decrease
exploration by driving exploitation. If exploration is needed and beneficial on a task this

suppression effect will produce a negative goal setting effect. If exploration is not

47

needed, and rather exploitation is beneficial, this suppression effect will produce a
positive goal setting effect. If exploration is beneficial but interrupted frequently enough
by dynamics of the task neither effect will gain enough temporal traction to establish
itself, and a neutral goal effect will result.

These findings are consistent, as shown earlier, with previous goal setting
research in that complexity is likely — in many studies - to be driven by these need for
exploration effects. If a task is simple exploration is likely to be detrimental to
performance — thus difficult specific goals that suppress exploration will drive the
numerous strong positive goal effects found on simple tasks. The more complex a task
is, the more likely it is that exploration will be needed for high performance. When
difficult specific goals are given in this situation, participants will still explore less than
participants given do your best goals, driving a reduction of the positive goal setting
effect. This reduction is again consistent with the findings of Wood, Mento, and Locke
(1987)

What cannot be explained by the findings of Wood, Mento, and Locke (1987) are
situations where negative goal setting effects arise, especially on simple tasks, or where
the relationship between complexity and the level of the goal setting effect operates
opposite to their predictions. The simple tasks where this negative goal setting effect
occurs tend to be insight based in some way or another and thus possess a need for
exploration. Huber’s (1985) maze task and Sweller and Levine’s (1982) maze task are
both good examples of this where such an effect is found. Both of these cases can be
explained using exploration as an explanation. Using this new theory as a template for

experimental design also allowed both of these cases, including the reversal of the

48

relationship of complexity to goal setting effects, to be illustrated empirically in this
experiment.

One question that should fall out of these results rests on the non-significant
findings in terms of relevant individual difference measures predicting exploration in the
task. Neither of the operationalizations of exploration was affected by any of the trait
measures: curiosity, learning goal orientation, and performance avoid goal orientation.
While Seijts and Latham (2001) showed that learning goals promote exploration, it was
an extension of this result to suggest that individual goal orientation may also drive this
effect. While it still may be the case that it does in weaker situations, it may be that the
situational effects in this case were simply too strong. This would also explain why
curiosity failed to predict exploration. It is possible that the trait measures taken were too
distal, and that a collection of state measures during the experiment may have had a more
proximal linkage to the behavioral measures and thus shown an effect. It is also true that
while the operationalizations of exploration are the best that can be achieved in this task
they are by no means perfect representations. More than anything it may be that the
situation presented is fairly strong and decreased the effects of distal trait measure
linkages to behavior.

Additionally, the worlds in this task were small enough that some participants
may have crossed from one side of the world to the other haphazardly and driven only by
chance. This was clearly not the case for all participants, as effects were found, but the
size of the world may have led to more noise in the data than if participants had
completed the task with substantially larger areas to explore and exploit. World 2 is

likely to have fared the worst in this situation, as participants who moved directly to the

49

rich resources and followed a orderly pattern of collection had a reasonable chance of
being able to collect all of the resources contained within. Faced with a now empty half
of the world they would likely cross to the other side not out of exploration but out of
lack of anything more to exploit. While this is problematic, it is not counter to the
results, as if this did happen it would only reduce the power of the effects that were
found, as people who were exploiting might inadvertently have higher exploration scores
while still (technically) exploiting. The fact that this would be more likely toward the
end of the task (as people become more skillful in collection) might actually explain the
diminishing linkage between exploration and performance late in the experiment in world
2. Thus, while it is not a detrimental sticking point of the current research, future work
on this task is likely to be benefited by the use of larger worlds.
Implications

Complexity - Unlike previous studies and meta-analyses, this study was able to
manipulate objective complexity in a way befitting Wood’s (1986) original mathematical
ideals. Due to the fact that only dynamic complexity was altered through the task design,
it is possible to claim that world 3 was in fact ordinally more complex than world 1.
While this only mapped to individual’s perceptions of the task in the pilot and was not
found on the larger population, it still has implications for task complexity theory in '
providing a fiamework by which to make manipulations to task complexity. As
complexity theory based on Wood (1986) currently stands, a scientific statement can only
be made about the relative standing of one task to another when only one variable of the
task (component, coordinative, or dynamic complexity) is changed, holding the other two

constant. Instead of making claims across widely disparate levels of tasks, more work

50

should be applied in understanding the effects of each of these different types of
complexity. Work on this level may eventually lead to the discovery of Wood’s (1986)
elusive a, B, and y.

Explorgtion and problem representation — The main implications to the area of
problem representation deal with the distinction of insight problems from other types of
tasks. Insight problems are unique in the way in which they need to be solved — a
solution must come from a distinctly different area than where one might start. For
example, a golfer may spend years and thousands of hours swinging golf clubs in order to
see small but steady improvements in performance. Conversely, a golfer may only need
to make one change of mindset to realize they have been holding the club upside-down,
and likely see immediate improvement in performance afier their change. The first is an
example of a problem solved by exploitation, the second a problem solved by
exploration. Providing a specific, difficult goal to the first individual may help drive their
performance, providing a specific, difficult goal to the second may simply frustrate them.

Insight problems in the real world are also likely to be without known solution.
These types of problems — in whatever form they may take —- may then be benefited by
using a ‘do your best’ goal approach as opposed to specific, difficult goals.

Exploration and goal setting theogg — In effect this entire paper is about the
implications of exploration as it applies to goal setting theory. If a task can be said to
require exploration, then specific, difficult goals should not be used in conjunction with it
— performance will be decreased. If a task can be said to require exploitation, then
specific difficult goals should be used in conjunction with it — performance will be

improved. This has been shown on a simple task, and with firture research on tasks of

51

medium and complex tasks it stands to reason that this advice can be applied without
worry of complexity, only with concern for need for exploration.

Earlier in the paper the idea of mastery and performance goals was discussed
briefly, somewhat dismissing mastery goals at the time in order to focus on performance
goals. Mastery goals, through their effects on learning, may be a proper substitute for ‘do
your best’ goal when specific, difficult (performance) goals are found to be detrimental
due to their effects on exploration. More research is needed that specifically links these
ideas to the processes of exploration while taking into account the amount of exploration
needed on the task.

The use of mastery goals comes with a caveat, as well; the use of specific,
difficult mastery goals may in the end have similar effects on unwarranted exploration
the same as specific, difficult performance goals. That is to say that while an
experimenter may be able to use a mastery goal to provoke a participant to explore parts
of their world that they otherwise might not find, this provocation may deter them from
other, more self-motivated and task-irrelevant exploration. This idea may not be
worrisome in a lab setting where the experimenter likely knows the best way for a task to
be completed, but in the real world, or with more complex tasks, any specific, difficult
goals — mastery or performance —may stop an individual from ‘thinking outside the box’,
so to speak.

Exploration ml reinforcement leappflg — The area of reinforcement learning has
already grappled with the ideas of exploration and exploitation to a large degree. One of
the largest implications for this area is that goals can seemingly be used as a trigger to

shift this balance. Almost impossible to predict (or perhaps imagine!) would be the

52

application of these ideas to reinforcement learning as told by automated agents, or
computers. Is it possible to model this behavior in a way that a computer can follow?
Are these drives of goals and curiosity innate to the human consciousness, or can they in
fact be adequately replicated? These questions are obviously far beyond the scope of this
paper, but answers to them may help to better understand just exactly how individuals are
being driven in terms of goals and exploration.

Other implication - The fact that goals were set at the 85th percentile is in proper
keeping with the goal setting literature. The fact that they were normed from the first
trial exclusively is less in keeping. While the reasons for this use are sound, it is easy to
imagine that using a specific, difficult goal from a later trial (and thus a higher score)
would have had a different effect on individuals. This is an important distinction, and
relies very heavily again on the distinction between insight and non-insight problems.
Goals were normed from the first trial in this study because it was desired that
participants strive to do the best they could do while exploiting. The insight to this
problem was that one could stop exploiting and explore. Participants fail to make this
jump because they are focused on their goals. In this experiment it was almost too easy
to find the rich resources; the problems of the size of the world have and will continue to
be discussed.

If the goal was set higher than a participant could achieve through exploitation,
they would be triggered that something was wrong. In this way, specific, difficult goals
set with full knowledge and control of the world may actually drive a behavior that
mimics exploration. It is argued that this would no longer be exploration — this would

simply be a continuation of exploitation.

53

l_’_rpctic2_rl implicatiog — When these ideas are removed from a lab setting a
number of questions and possible implications arise. First and foremost, as stated earlier
in this paper, using the incorrect balance of exploration and exploitation in the real world
can, at its absolute worst, result in death of a species. At the corporate level it can result
in bankruptcy or dissolution. At the very least it can lead to outcomes more unfavorable
than were met in the lab by undergraduate psychology students who were simply having
firn playing a game. It is likely that exploration in the real world holds a larger risk than
could be replicated in a lab. Despite this, it is suggested that the above findings should
still hold, possibly with a number of qualifications to be uncovered by future research.
One idea is that while the specific, difficult goals were used in this study were strictly
outcome based, similar results may be found when specific, difficult time goals are
placed on individuals or organizations. It may also be the case that many of the problems
faced by organizations as a whole can be conceptualized as insight problems — problems
which no one may have a solid answer to. Simply making slide rulers 100 times faster at
100 times the efficiency might stop the slow onslaught of the electronic calculator, but it
will never stop it. If exploration is needed, limit the use of specific, difficult goals. If
exploitation is needed, use specific, difficult goals.

Limitations

This study was completed on college students in a psychology subject pool, on a
specific task, and as such it may be the case that future replications on diverse
populations are needed to show full generalizability. That said, many of the results of the
goal setting literature have first been found in college populations participating in lab

settings, and some replication has also been shown on a different task in this study (pilot

54

1) as well as in a different situation (exploratory analysis on those returning to the
dispersed resources in world 3). It would be useful to replicate the results of this study on
a task of both medium and high complexity in order to show results across the full
spectrum of task complexity. While the task used in this study was a relatively simple
one, results fitting with the hypotheses proposed have already been found on a wide
spectrum of tasks (Burns & Vollmeyer, 2002; Huber, 1985; lvancevich, 1974; Pritchard,
Jones, Roth, Stuebing, & Ekeberg, 1988; Sweller & Levine, 1982). What is missing in
these studies are measures of exploration both needed and completed on the task and the
effects of goals on these measures — the foundation for which has been put forth in this
study. The results of this study should be generalizable to any number of tasks as long as
the amount of exploration needed on the task can be qualified. If exploration is needed,
specific difficult goals will hurt, and if exploration is not needed, specific difficult goals
will help.

In terms of limitations of the task itself, and as already mentioned, it might also
have been possible that the size of the world in which participants acted imposed
constraints on the total amount of exploration that participants could engage in. This
might also have produced a constraint from the opposite end, as those exploiting may
have eventually gotten good enough at the task to run out of things to exploit, prompting
exploration.

The orientation of the world as well, with participants being able to find three
‘walls’ of the world fairly quickly may have prompted them to find the final wall through
some other motivational mechanism. Individual differences such as need for cognitive

closure or tolerance of ambiguity may well have predicted some of this behavior, and

55

future work on this task should either address this issue by ‘removing’ the walls or
examining these processes.
Next Steps

As the task used is in its early stages, much was learned though this study which
may be useful both in development of the task as well as redesign to look at additional
effects and relationships.

By increasing the size of the world, and thereby pushing the walls of the lower
half of the world away from the participant, participant behavior may be drastically
altered. Predominantly, this would open up whole new areas for them to explore, instead
of — in effect — ‘funneling’ them toward the top of the screen. This sort of world could be
useful in manipulating the benefits of exploration by giving a larger area of space that
could be made beneficial or detrimental via exploration. Though this, participants could
be given a wider range of possible behaviors. While this may complicate scenario design
it allows the task to address a wider range of questions, such as how much participants
will continue to devote to exploration that may not be working (e. g. sunk cost
paradigms).

Exploration was also examined to a large extent in this study, but exploitation was
not concretely operationalized (other than as the absence of some exploration). There are
a number of ways that exploitation can potentially be operationalized in future work. The
number of dispersed pools that they visit or even the redundancy between pools they visit
from one trial to the next could both be viable ways at developing a variable to

incorporate exploitation into firture work and analysis.

56

There is also some potential need to show that goals created in different ways
(e. g. proximal goals normed at each trial, distal terminal goals from final trials) have
similar effects. While the world structures of world 1 and world 3 may need to be
changed a bit in order to accommodate this change, world 2 is the place where goals
seem to have the most leeway in their location on the score scale. Goals in world 2 have
the potential to be raised by almost 40 points before hitting the exploitation ceiling of 180
points. This task, and this design of the task possesses an interesting point on the score
scale — this exploitation ceiling that is the maximum that can be scored without exploring.
This point on the scale may be a cut point where effects can be further changed. Future
research on larger versions of world 1 or 3, or world 2 with different levels of goals may
give more insight into this unique property.
Conclusions

Negative goal setting effects, as well as situations in which the relationship
between complexity and goal effects operate contrary to the predictions of Wood, Mento,
and Locke (1987), cannot be explained under current theories with the exception of
dismissing them as experimental flukes. This paper has shown that these effects are not
flukes, but situations in which the need for exploration on the task does not follow the
simple ‘more complex tasks need more exploration’ relationship. Complexity works
sometimes — when the above statement holds true — but exploration is a deeper and more
encompassing explanation of effects. While this paper was aimed at explaining these
‘flukes’ of goal setting, even positive goal effects are explainable by exploration,

allowing exploration as an explanation of goal setting effects to account for all the types

57

of empirical results in the goal setting literature, not just those that fit preconceived

notions.

58

Appendixes

Table 1: Regression of cumulative exploration on performance over time by world.

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

World Trial Number R2 B SE B Standardized j}
1 .16 108.38 25.04 .40**
1 2 .27 117.88 19.61 .52**
_ 3 .32 150.55 21.85 .57M
(N400) 4 .33 162.69 23.43 .58M
5 .34 179.73 25.19 .59**
1 .70 -116.56 8.04 -.84**
2 2 .18 -70.97 15.92 -.42**
_ 3 .11 -62.89 19.27 -.32**
(N412) 4 .03 -3337 20.72 -.17
5 .04 47.25 23.06 -.21*
1 .33 119.86 17.63 .57M
3 2 .26 99.51 16.95 .51**
13*
01:98) 3 .37 143.73 19.19 .61
4 .29 151.67 24.18 .54“
5 .13 126.67 33.82 .36"

 

 

 

 

 

 

 

 

 

 

Note: * - Significant at the p<.05 level; ** - significant at the p<.01 level.

59

 

Table 2: Linear regressions of proposed moderators on exploration: results for hypothesis

3, 4a, and 4b (5x1); (N=310).

 

Predictor

 

Trait Curiosity Trait Learning Goal Trait Performance

Orientation . . . a
Avord Orrentatron

 

B SEBB B SEBB B SEBB

 

EXl Triall .002 .017 .008 .011 .018 .037 -.023 .017 -.112

 

EXlTria12 -.002 .019 -.007 .017 .020 .052 -.017 .020 -.072

 

EXl Trial3 .005 .019 .018 .022 .020 .070 -.021 .020 -.089

 

EXl Trial4 -.004 .019 -.012 .020 .020 .064 -.025 .021 -.102

 

 

EXl TrialS -.007 .018 -.024 .016 .019 .051 -.018 .020 -.076

 

 

 

 

 

 

 

 

 

 

 

a ’ . . . . . . . .
Note: Perfonnance-avord onentatron was only tested on condrtrons 1n which specrfic

goals were present (N=158). Tests on full sample (N=310) were consistent with these
results.

* Significant at the p <.05 level.

60

Table 3: Linear regressions of proposed moderators on exploration: results for hypothesis

3, 4a, and 4b (sz).

 

Predictor

 

Trait Curiosity Trait Learning Goal Trait Performance

Orientation . . . a
Avord Orrentatron

 

B SEBB B SEBB B SEBB

 

EX2 Triall .077 .055 .086 .054 .058 .058 -.083 .057 -.124

 

EX2Tria12 -.016 .058 -.017 .046 .061 .046 .017 .062 .024

 

EX2 Trial3 -.031 .058 -.033 .012 .061 .012 -.051 .062 -.070

 

EX2 Trial4 .069 .057 .075 .035 .060 .036 -.032 .064 -.043

 

 

EX2Tria15 -.015 .056 -.017 -.019 .059 -.019 .009 .063 .013

 

 

 

 

 

 

 

 

 

 

 

a . . . . . . . .
Note: Performance—avord onentatron was only tested on condrtrons 1n which specrfic

goals were present (N= 158). Tests on full sample (N= 310) were consistent with these
results.

* Significant at the p <.05 level.

61

 

Table 4: Results of repeated measures analysis on performance and exploration over time

 

 

 

 

 

 

 

 

 

 

 

by goal condition and world.
Dependant Variables

Result Numerator Denominator Performance 3 Exploration Exploration

df df (EXl)a (EX2)a
Time 4 1216 123.05“ 426.5“ 37.65“
World 2 304 30.57M 5.46M 25.41M
Goals 1 304 .07 . 4.17* 20.68"
Time by 8 1216 3.45“ 10.62“ 17.04"
World
Time by 4 1216 3.84“ 4.21M 289*
Goals
World 2 304 9.02“ 2.21 8.98"
by Goals
Time by 8 1216 3.01“ 1.48 1.06
World
by Goals

 

 

 

 

 

 

*Significant at the p<.05 level.

"Significant at the p<.01 level.

a — F values.

62

Figure 1 - 'Do your best' goals

 

0.80—

Probability
3.

9

a.

o
I

0.20“

 

 

 

Illlllllllllllllllll
12 3 4 5 6 7 e 91011121314151617181920
time

arm

—ann1
—-arm2

arm3
—-amr4

Figure 2 - Specific, difficult goals

 

0.80-

p
8
l

Probability

0.40-

0.20—

 

I: R
‘

e/b . 3 ‘
_ t"\ e 2
a ‘2-2-2 2‘. = .. »

r. 3‘:

 

 

llllllllllll
91011121314151617181920
time

‘—
~—
g.)—
#
aa—
a,—
N—
m—

64

arm

—ann1
—arm2

arm3
—arm4

Figure 3 - Performance over time

 

30‘

25‘

N
‘?

Performance
3.

5—

 

o-I

 

 

 

11 [11111111111111]
12 3 4 5 6 7 e 91011121314151617181920
time

65

condition

— DYB goals
— SD goals

Figure 4 —— World 1

M. ‘—-.’~ .r—1 “1“ vaqxe 0‘ ~Mm. - , ”WA “P—I‘f?—' 1

 

66

Figure 5 — World 2

 

67

Figure 6 — World 3

. aw: V“: "vary”? {-1171 ‘ 'fzz’v #4.: WM’:

 

68

Figure 7 — Start of Trial (zoomed out)

swummee
mamaqum
EEEEEHE

gsmfimae

 

69

Figure 8 — After some exploring in trial (zoomed out).

 

70

REFERENCES

Berry, D.A., & Fristedt, B. (1985).Bandit problems: Sequential allocation of experiments.
New York, NY: Chapman and Hall.

Bird, R., Smith, E., & Bird, D. (1999). The hunting handicap: costly signaling in human
foraging strategies. Behavioral Ecology and Sociobiology, 50, 9-19.

Burns, B.D. & Vollmeyer, R. (2002). Goal specificity effects on hypothesis testing in
problem solving. The Quarterly Journal of Experimental Psychology, 55, 241-
261.

Campbell, D. (1988). Task complexity: a review and analysis. Academy of Management
Review, 13, 40-52.

Carr, H. A. Psychology: A study of mental activity. New York: Longrnans, Green, 1925.

Cronicle, E.P., MacGregor, J .N., & Ormerod, TC. (2004). What makes an insight
problem? The roles of heuristics, goal conception, and solution recoding in
knowledge-lean problems. Journal of Experimental Psychology: Learning,
Memory, and Cognition, 30, 14-71.

Das, S. & Kamenica, E., (2005). Two-sided bandits and the dating market. Proceddings
of the Nineteenth International Joint Conference on Artificial Intelligence, 1 9,
947-952.

DeJong, K.A. & Spears, W.M., (1993). Using genetic algorithms to solve NP-complete
problems. Proceedings of the 3rd International Conference on Genetic
Algorithms, 124.

DeShon, R.P. & Alexander, RA. (1996). Goal setting effects on implicit and explicit
learning of complex tasks. Organizational behavior and human decision
processes, 65, 18-36.

Earley, P.C., Connolly, T. & Ekegren, G. (1989). Goals, strategy development, and task
performance: some limits on the efficacy of goal setting. Journal of Applied
Psychology, 74, 24-33.

Earley, P.C., Connolly, T., & Lee, C. (1989). Task strategy interventions in goal setting:

the importance of search in strategy development. Journal of Management, 15,
589-602.

Fowler, H., (1965). Curiosity and exploratory behavior. New York: Macmillan.

Garcia, Rosanna, Calantone, Roger and Levine, Ralph (2003). The Role of Knowledge in

71

Resource Allocation to Exploration versus Exploitation in Technologically
Oriented Organization. Decision Science, 34, 323—349.

Harlow, HF. (1950). Learning and satiation of response in intrinsically motivated
complex puzzle performance by monkeys. Journal of Comparative Psychological
Physiology, 43, 289-294.

Houghton Mifflin Company. (2006) The American Heritage® Dictionary of the English
Language, Fourth Edition. Houghton Mifflin

Huber, V.L. (1985). Effects on task difficulty, goal setting, and strategy on performance
of a heuristic task. Journal of Applied Psychology, 70, 492-504. .

lvancevich, J .M., (1974). Changes in performance in a management by objectives
program. Administrative Science Quarterly, 19, 563-574.

James, W. (1890). Principles of psychology. New York: Holt.

Jervis, R. (1988). Perception and misperception in international politics. Princeton,
N.J.: Princeton University Press.

Kaelbling, L.P., Littman, M.L., & Moore, A.W. (1996). Reinforcement learning: a
survey. Journal of Artificial Intelligence Research, 4, 237-285.

Kanfer, R. & Ackerman, P. (1989). Motivation and cognitive abilities: an
integrative/aptitude-treatment interaction approach to skill acquisition. Journal of
Applied Psychology, 74, 657-690.

Kashdan, T.B., Rose, P., & F incham, FD. (2004). Curiosity and exploration: facilitating
positive subjective experiences and personal growth opportunities. Journal of
Personality Assessment, 82, 291-305.

Klein, H.J, Wesson, M.J., Hollenbeck, J .R., Wright, P.M., & DeShon, RP. (2001). The
assessment of goal commitment: A measurement model meta-analysis.
Organizational Behavior and the Human Decision Processes, 85, 32-55.

Kozlowski, S.W.J & Bell, BS. (2006). Disentangling achievement orientation and goal
setting: Effects on self-regulatory processes. Journal of Applied Psychology, 91,
900-916

Latham, G.P. & Klinne, SB, (1974). Improving job performance through training in
goal setting. Journal of Applied Psychology, 59, 187-191.

Latham, G.P & Seijts, GB. (1999). The effects of proximal and distal goals on

performance on a moderately complex task. Journal of Organizational Behavior,
20, 421 -429.

72

Litrnan, J ., Hutchins, T., & Russon, R., (2005). Epistemic curiosity, feeling-of-knowing,
and exploratory behavior. Cognition & Emotion, 19, 559-582.

Litrnan, J .A. & Spielberger, CD. (2003). Measuring epistemic curiosity and its diversive
and specific components. Journal of Personality Assessment, 80, 75-86.

Locke, E.A. & Latham, G.P., (1990). A theory of goal-setting and task performance.
Englewood Cliffs, NJ: Prentice-Hall.

Locke, E.A. & Latham, GP. (2002). Building a practically useful theory of goal setting
and task motivation. American Psychologist, 5 7, 705-717.

Locke, EA. (1968). Toward a theory of task motivation and incentives. Organizational
Behavior and Human Performance, 3, 157—189.

Locke, E.A., Shaw, K.N., Saari, L.M., & Latham, GP. (1981). Goal-setting and task
performance: 1969-1980. Psychological Bulletin, 90, 125-152.

Loewenstein, G. (1994). The psychology of curiosity: a review and reinterpretation.
Psychological Bulletin, 116, 75-98.

March, J .G. (1991). Exploration and exploitation in organizational learning.
Organizational Science, 2, 71-87.

Miller, C.S., Lehman, J .F ., & Koedinger, KR. (1999). Goals and learning in
microworlds. Cognitive Science, 23, 305-336.

Nooteboom, B. (1999). Discovery and organization: priorities in the theory of
innovation. Paper for the 1999 DR UID conference.

Pritchard, D., S.D.Jones, P.L.Roth, K.K.Stuebing, & S.E.Ekeberg (1988). Effects of
group feedback, goal setting, and incentives on organizational productivity.

Journal of Applied Psychology, 73, 337-358.

Pyke, CH. (1984). Optimal foraging: a critical review. Annual Review of Ecology and
Systematics, 15, 523-575.

Random House, Inc. (2006). Random House Unabridged Dictionary. Random House.

Reio, T.G., Petrosko, J .M., Wiswell, A.K., & Thongsukmag, J. (2006). The measurement
and conceptualization of curiosity. The Journal of Genetic Psychology, 1 6 7, 117-
135.

Rode, C., Cosmides, L, Hell, W. & Tooby, J. (2001). When and why do people avoid

73

unknown probabilities in decisions under uncertainty? Testing some predictions
fiom optimal foraging theory. Cognition, 72, 269-304

Rothaermel, F.T. & Deeds, D.L. (2004). Exploration and exploitation alliances in
biotechnology: a system of new product development. Strategic Management
Journal 25, 201—221.

Seijts, G.H. & Latham, GP. (2001). The effect of distal learning, outcome, and proximal
goals on a moderately complex task. Journal of Organizational Behavior, 22,
291-307.

Seijts, G.H. & Latham, GP. (2005). Learning versus performance goals: When should
each be used? Academy of Management Executive, 19, 124-131.

Seijts, G.H., Latham, G.P, Tasa, K., & Latham, B.W. (2004). Goal setting and goal
orientation: an integration of two different yet related literatures. Academy of
Management Journal, 47, 227-239.

Sweller, J. & Levine, M. (1982). Effects of goal specificity on means-ends analysis and
learning. Journal of Experimental Psychology, 88, 463-474.

VandeWalle, D. (1997). Development and validation of a work domain goal orientation
instrument. Educational and Psychological Measurement, 5 7, 995—1015.

Vollmeyer, R., Burns, B., & Holyoak, K. (1996). The impact of goal specificity on
strategy use and the acquisition of problem structure. Cognitive Science, 20, 75-
100.

Wood, R. (1986). Task complexity: definition of the construct. Organizational Behavior
and Human Decision Processes, 3 7, 60-82.

Wood, R., Mento, A., & Locke, EA. (1987). Task complexity as a moderator of goal
effects: a meta-analysis. Journal of Applied Psychology, 72, 416-425.

74

IIIIIIIIIIIII

293 03062 7990

III

31

 

IIIIIIIIIII