Learning and Decision-Making in Active Cognition
Learning and decision making is one of the universal cornerstones of human and animal life. There are several situations where these processes are passive or secondary, but there are several others where the individual can actively control where to gain information to learn more effectively. For example, selecting which restaurant to visit next with the aim of finding out a consistently good dining option. Even though there have been some efforts to model problems in active cognition, a large class of such models is algorithmic, and only concerns itself with how such behavior may be replicated, but not how and why the algorithm actually comes about. This leads to a dearth of computational models that can explain the behavior in a principled manner starting from basic, justifiable assumptions.
Employing tools such as hierarchical Bayesian modeling and stochastic control, this dissertation proposes novel models for several problems in cognition with the goals of enriching our understanding of how the brain might tackle such problems, as well as to provide useful guidelines for building artificial agents that could rival human performance on such problems. More specifically, we investigate three related problems and propose Bayesian models for each of these. Active sensing problem, which deals with optimally allocating sensory resources in a given sensing task. Competitive foraging problem, which deals with understanding foraging in an uncertain environment and in the presence of competing agents who affect each other's reward. And lastly, multi-attribute decision-making problem, which deals with understanding how humans make choices between multi-attribute alternatives and how such a choice mechanism could lead to the well-studied contextual effects.