Agent Influence and Intelligent Approximation in Multiagent Problems

Dec-POMDPs are worst-case intractable. Recent approximation algorithms have achieved positive results, but performance depends upon parameters set in advance, and little is known about how to choose those settings. We provide an information-theoretic measure of agent influence, shown to be (1) a good indicator of algorithm performance, and (2) a clue to setting algorithm parameters, potentially improving runtime and memory requirements dramatically without sacrificing quality.