Multiagent reinforcement learning using function approximation


Abul O., Polat F., Alhajj R.

IEEE TRANSACTIONS ON SYSTEMS MAN AND CYBERNETICS PART C-APPLICATIONS AND REVIEWS, cilt.30, sa.4, ss.485-497, 2000 (SCI-Expanded) identifier identifier

Özet

Learning in a partially observable and nonstationary environment is still one of the challenging problems In the area of multiagent (MA) learning. Reinforcement learning is a generic method that suits the needs of MA learning in many aspects. This paper presents two new multiagent based domain independent coordination mechanisms for reinforcement learning; multiple agents do not require explicit communication among themselves to learn coordinated behavior. The first coordination mechanism Is perceptual coordination mechanism, where other agents are included in state descriptions and coordination information is Learned from state transitions. The second is observing coordination mechanism, which also includes other agents in state descriptions and additionally the rewards of nearby agents are observed from the environment. The observed rewards and agent's own reward are used to construct an optimal policy. This way, the latter mechanism tends to increase region-wide joint rewards. The selected experimented domain is adversarial food-collecting world (AFCW), which can be configured both as single and multiagent environments, Function approximation and generalization techniques are used because of the huge state space. Experimental results show the effectiveness of these mechanisms.