Behavior control of multi-robot using the prior knowledge based reinforcement learning
Meiping Song, Guochang Gu, Rubo Zhang · 2004
In the partially known environment, it was hard to control the robot's behavior exactly and flexibly. The rule-based method can't cover all the possible conditions, and the traditional reinforcement learning method also has the problem of convergence. The prior-knowledge based reinforcement learning prompted here combines the advantages of these two methods and avoids the above disadvantages. It takes the determinately known rules as prior-knowledge to train the learner, so as to guarantee the direction and convergence of learning and speed up the learning. At the same time, the adaptive quality of learner makes it automatically exploit the unknown environment. This makes up the shortcoming of the incompletely known rules. When this method is applied to the action integration of robot's behavior control in the pursuit-evasion game, it overcomes the toothed problem of rule-based control and the unexpected cases of traditional reinforcement learning. The robot is proved experimentally to circumambulate the obstacles smoothly, and collide with the obstacle rarely.