Bounded-Rational Decision-Making Hierarchical Models for Autonomous Agents
Abstract
Empirical evidence suggest that humans do not always act as perfectly rational agents when making decisions. Instead of unconditionally maximizing an internal utility function, instead they tend to maximize this function subject to computational, information, or energy constraints. According to thisbounded rationality paradigm, human beings tend to weight the benefits of reaching the ???best??? or ???most informed??? decision, versus the pitfalls associated with the time, energy and effort needed to collect all the information to reach that decision. A bounded rational decision-maker therefore needs to strike a balance between performance (encoded in terms of desired utility) and available resource costs.The main objective in the proposed research is to develop decision-making models and algorithms for human-like decision-making for demanding tasks in the presence of uncertainty and/or adversarial agents that require significant cognitive effort under time-constraints, when exact optimality/rationalitymay be elusive. The proposed approach can be used to also better explain how humans make decisions and can therefore be used to establish trust in human/machine teams by generating predictable behaviors. The bounded rational model is highly relevant for the development of the next generation ofautonomous systems, having (expert) human-like perception and decision-making capabilities, especially when these autonomous agents need to interact with humans. The interaction between autonomous robotic agents (or between intelligent agents and humans) will be investigated while considering the decision-making constraints of the agents. Instead of focusing on maximizing expected utility (as in standard MDP formulations) the proposed approach will be based on maximizing expected free utility (or free energy), a quantity that encapsulates in a precise manner the effect of information and computational constraints of the decision-making process. The results will be applied to the case of multiple (perhaps conflicting) agents in the presence of uncertainty. To avoid the infinite regress that hinders the computation of (purely rational) Nash equilibria, we will adopt a bounded rationality framework that either investigates correlated equilibria or restricts each player???s depth of recursive inference. The proposed framework also intrinsically generates a hierarchy of abstractions for decision making brought about by the severity of information-processing costs, with each abstraction tailored to the ???right??? level of granularity (that is, the level that is commensurate to the available computational resources). The theory and methodologies developed in this research will make it possible to run highlysophisticated algorithms inside the ???brain??? of the future autonomous systems of interest to the Navy, thus increasing their reliability, predictability, and fail-safe operation. The proposed research will also have an immediate impact on adversarial interaction between two teams of agents of variable rationality, a common scenario in most military encounters.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jul 10, 2018
- Source ID
- N000141812375
Entities
People
- Panagiotis Tsiotras
Organizations
- Georgia Tech Research Corporation
- Office of Naval Research
- United States Navy