AILo Talk - Rafael Cunha, University of Groningen

Title: Compounding behaviours for fast multi-agent reinforcement learning: The one-agent-at-a-time approach

Abstract:

Transfer learning has emerged as a promising avenue in reinforcement learning (RL). By leveraging prior experiences and policies, agents can efficiently adapt to new tasks. One way to achieve this is to use the "composition of behaviors" via a method termed 'successor features.'

However, the introduction of multiple agents in the RL landscape complicates matters. The exponential growth in action possibilities demands more refined strategies for efficient learning. One way to circumvent this is by using the "one-agent-at-a-time" approach, a methodology that streamlines the search process in multi-agent scenarios by sequentially focusing on individual agents, albeit with an expanded state consideration.

Our contribution is a novel theorem, specifically crafted for transfer learning in multi-agent contexts. Building on the one-agent-at-a-time framework, this theorem ensures a linearly growing search space relative to the agent count, guaranteeing that each newly derived policy will not underperform its predecessor. This work paves the way for more efficient strategies for transfer learning in multi-agent RL.