Learning to Coordinate without Sharing Information

Researchers in the eld of Distributed Arti cial Intelligence DAI have been developing e cient mechanisms to coordinate the activities of multi ple autonomous agents The need for coordina tion arises because agents have to share resources and expertise required to achieve their goals Previous work in the area includes using sophis ticated information exchange protocols investi gating heuristics for negotiation and developing formal models of possibilities of con ict and co operation among agent interests In order to han dle the changing requirements of continuous and dynamic environments we propose learning as a means to provide additional possibilities for e ec tive coordination We use reinforcement learning techniques on a block pushing problem to show that agents can learn complimentary policies to follow a desired path without any knowledge about each other We theoretically analyze and experimentally verify the e ects of learning rate on system convergence and demonstrate bene ts of using learned coordination knowledge on simi lar problems Reinforcement learning based coor dination can be achieved in both cooperative and non cooperative domains and in domains with noisy communication channels and other stochas tic characteristics that present a formidable chal lenge to using other coordination schemes