Multi-Agent Reinforcement Learning and Adaptive Neural Networks.
Final rept. 1 Apr 93-30 Apr 96,
MASSACHUSETTS UNIV AMHERST DEPT OF COMPUTER SCIENCE
Pagination or Media Count:
This project investigated learning systems consisting of multiple interacting controllers, or agents each of which employed a modern reinforcement learning method. The objective was to study the utility of reinforcement learning as an approach to complex decentralized control problems. The major accomplishment was a detailed study of multi-agent reinforcement learning applied to a large-scale decentralized stochastic control problem. This study included a very successful demonstration that a multi-agent reinforcement learning system using neural networks could learn high-performance dispatching of multiple elevator cars in a simulated multi-story building. This problem is representative of very large-scale dynamic optimization problems of practical importance that are intractable for standard methods. The performance achieved by the distributed elevator controller surpassed that of the best of the elevator control algorithms accessible in the literature, showing that reinforcement learning can be a useful approach to difficult decentralized control problems. Additional empirical results demonstrated the performance of reinforcement learning-systems in the setting of nonzero-sum games, with mixed results. Some progress was also made in improving theoretical understanding of multi-agent reinforcement learning.