Recognizing environmental change through multiplex reinforcement learning in group robot system
Toshio Fukuda, D. Funato, Fumihito Arai · 2003
In recent years, reinforcement learning has been used as one of methods for making decisions for autonomous robots. However, the algorithm for learning concludes a parameter, and has a fault that adaptation speed for unknown environment changes, according to the value of the parameter. Because of the fault, the robot with reinforcement learning is often delayed in adapting to environmental change. In this paper, we propose a method to reduce the adaptation delay. Our idea is to make it possible for the robot to know if environmental change has happened. We found that if the robot uses reinforcement learning with two different parameters, the correlation between the reinforcement values lowers due to environmental change. Thus, we use the correlation as an index showing environmental change. The effect is proved through computer simulation.