eduzhai > Applied Sciences > Engineering >

Emergent cooperation through mutual information maximization

  • Save

... pages left unread,continue reading

Document pages: 20 pages

Abstract: With artificial intelligence systems becoming ubiquitous in our society, itsdesigners will soon have to start to consider its social dimension, as many ofthese systems will have to interact among them to work efficiently. With thisin mind, we propose a decentralized deep reinforcement learning algorithm forthe design of cooperative multi-agent systems. The algorithm is based on thehypothesis that highly correlated actions are a feature of cooperative systems,and hence, we propose the insertion of an auxiliary objective of maximizationof the mutual information between the actions of agents in the learningproblem. Our system is applied to a social dilemma, a problem whose optimalsolution requires that agents cooperate to maximize a macroscopic performancefunction despite the divergent individual objectives of each agent. Bycomparing the performance of the proposed system to a system without theauxiliary objective, we conclude that the maximization of mutual informationamong agents promotes the emergence of cooperation in social dilemmas.

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×