Individual Decision-Making in Uncertain and Large-Scale Multi-Agent Environments

摘要

Research undertaken in this initial performance period developed: (a) the first set of generally applicable approximation methods for the finitely nested interactive POMDP (I-POMDP) framework, and (b) novel probabilistic graphical models called interactive dynamic influence diagrams (I- DIDs) that generalize the well-known DIDs to multiagent settings. These methods provide approximation techniques for decision making in complex multiagent settings in reduced time and space facilitating scalability. Experiments reveal that the approaches generate solutions of flexible quality proportional to the computational resources allocated. In the context of human decision making, this research showed that a strategic setting that was relatively simple, realistic and competitive increased the tendency in subject to attribute higher levels of reasoning to others, which are consistent with typical levels of adversaries' reasoning.