MACC'97 Poster session
To investigate the potentials and limitations of multi-agent reinforcement learning, many attempts have been reported to let multiple reinforcement-learning agents synthesize coordinated decision policies needed to accomplish their common goals effectively. In this paper, we consider the Simulated Dodgeball(SD) as a multi-agent learning problem. In SD, multiple attacker agents are required to hit a single dodger agent by a ball. The dodger's behavior is manually-programmed in advance and it does not change, but the attackers are allowed to improve their own behavior. We implement each attacker agent by modular reinforcement-learning architecture and show how successfully the learning attacker agents synthesize effective coordinated behavior.
Multi-agent Systems, Multi-agent Learning, Reinforcement Learning, Modular Q-learning, Collective Behavior
PS
file(+gzip) (in Japanese)