Evolution of decision-making systems
Adaptive biological or engineered systems are adaptive because they can make decisions. Some systems such as viruses use their molecular composition – genetic information – to decide when to become lysogenic (dormant) or lytic (active). Others, such as self-driving cars, must use spatiotemporal information about obstacles, speed, and previous signs to determine when to turn or begin braking. Computational models of systems allow us to both engineer better systems, and create better scientific understanding about the dynamic world. The practice of modeling decision-making started with the study of interactions between rational agents on the spectrum of conflict and cooperation began with Von Neumann and Morgenstern's Theory of Game and Economic Behavior.Scenarios, called "games", are models designed and studied to increase understanding of conflict and cooperation between these agents. The games discussed here are Prisoner's Dilemma and Volunteer's Dilemma. Modern methods of analysis for games involving populations of interacting agents fail to predict the final strategy distribution among all agents. In chapter 2 I develop a new computational agent-based simulation used as an inductive study system to compare the deductive predictive capabilities of an analytical model that is capable of predicting the final distribution under idealized conditions. Lastly, I show a novel finding that the agent-based model suggests probabilistic, or mixed, strategies (such as probabilistic gene expression) are a result of the development and maintenance of cooperation in Volunteer's Dilemma.Game theory fails to provide tractable models for more complex decision-making situations, such as those with complex spatial or temporal dimensions. In these cases an algorithm of conditional logic may be used to simulate decision-making behavior. Yet still there are systems for which the use of an algorithm as a model is inadequate due to incomplete knowledge of the system. Perhaps the model makes too many generalizations, is limited by atomic discretization, or is otherwise incomplete. In these cases it is useful to compensate for deficits by using probabilistic logic. That is, we assume that a stochastic process can roughly describe those subprocesses not fully modeled.Lastly, algorithms as decision strategies can incorporate temporal information in the decision-making process. There are two ways temporal information can be used in an individual's conditional logic: evolutionary, and lifetime. The evolutionary approach has proved much more flexible as a means to discover and tune models of unknown decision-making processes. Neuroevolution is a machine learning method that uses evolutionary algorithms to train artificial neural networks as models of decision-making systems. There is currently a wide diversity of methods for neuroevolution that all share common structures of the types of problems being solved: those generally being cognitive tasks. Toward this end it would be useful if there were some properties common to all cognitive systems that could be incorporated into the optimizing objective function in order to enhance or simplify the evolutionary process. In chapter 3 and 4 I explore new methods of improving model discovery through neuroevolution and discuss the applicability of these methods for probabilistic models.
Read
- In Collections
-
Electronic Theses & Dissertations
- Copyright Status
- Attribution 4.0 International
- Material Type
-
Theses
- Authors
-
Schossau, Jorden D.
- Thesis Advisors
-
Adami, Christoph
- Committee Members
-
Punch, William
Ofria, Charles
Hintze, Arend
- Date
- 2017
- Program of Study
-
Computer Science - Doctor of Philosophy
- Degree Level
-
Doctoral
- Language
-
English
- Pages
- xiv, 97 pages
- ISBN
-
9781369762631
1369762631
- Permalink
- https://doi.org/doi:10.25335/xecb-8a28