eduzhai > Applied Sciences > Engineering >

Scalable Multi-Agent Reinforcement Learning for Networked Systems with Average Reward

  • Save

... pages left unread,continue reading

Document pages: 33 pages

Abstract: It has long been recognized that multi-agent reinforcement learning (MARL)faces significant scalability issues due to the fact that the size of the stateand action spaces are exponentially large in the number of agents. In thispaper, we identify a rich class of networked MARL problems where the modelexhibits a local dependence structure that allows it to be solved in a scalablemanner. Specifically, we propose a Scalable Actor-Critic (SAC) method that canlearn a near optimal localized policy for optimizing the average reward withcomplexity scaling with the state-action space size of local neighborhoods, asopposed to the entire network. Our result centers around identifying andexploiting an exponential decay property that ensures the effect of agents oneach other decays exponentially fast in their graph distance.

Please select stars to rate!


0 comments Sign in to leave a comment.

    Data loading, please wait...