Influence Based Reward Shaping Without a Heuristic
Keywords: multiagent systems, reward shaping, social influence
TL;DR: We borrow ideas from social influence to see if we can determine that one agent is "influencing" another teammate without using any domain knowledge.
Abstract: Learning based approaches have been used to coordinate multiagent systems in a variety of settings. A key challenge in multiagent learning is that agents are rewarded as a team, making it difficult to determine which agents took helpful actions. Influence based reward shaping helps address this by shaping each agent's reward signal to include credit for their actions as well as the actions of teammates they influenced. However, this requires a domain-specific definition of influence to determine who is influencing who, which may not be straightforward to define. This work-in-progress takes a step towards developing a domain-agnostic approach using causal influence. Our preliminary results show that we can tease out how one agent's actions affect its teammate's actions by simulating counterfactual scenarios.
Type Of Paper: Work-in-progress paper (max page 6)
Anonymous Submission: Anonymized submission.
Submission Number: 31
Loading