Discovering Optimal Strategies for Bounded Agents
Abstract
Making decisions in real environments requires effective use of limited computational resources and limited time. Humans constantly encounter these constraints, but nonetheless are still capable of outperforming automated systems in many contexts. The proposed research aims to explore how people identify effective cognitive strategies for decision-making, problem-solving and reasoning, and to translate the resulting discoveries into methods that can be used to improve automated systems. There are three objectives. First, expressing the problem of formulating strategies in terms of sequential decision problems, making it possible to leverage existing methods for reinforcement learning and planning. Second, evaluating proposed schemes for learning high-level actions developed in hierarchical reinforcement learning against human behavior, improving on these methods, and using the results to learn simple cognitive strategies. Third, exploring how these methods can be extended to learn algorithms that can instantiate more complex cognitive strategies. In addition to comparing performance to human behavior in a series of online experiments, the methods that are developed will be evaluated on challenge problems that tax the limits of human cognition. The results will provide insight into how people form cognitive strategies, as well as understanding of when human decision-making is likely to fail, and provide new tools that can be used to improve the performance of automated systems and better support human-machine interaction.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Apr 09, 2018
- Source ID
- FA95501810077
Entities
People
- Thomas L. Griffiths
Organizations
- Air Force Office of Scientific Research
- United States Air Force
- University of California Regents