Acceleration of game learning with reward propagation in segmented state space
暂无分享,去创建一个
Human can describe the world in discrete representations by recognizing characteristic phenomena in addition to the continuous one. It is thought that the discrete representation enables efficient problem solving. In this paper, we propose a method for finding a passing point which is important for reaching the goal by propagating the obtained reward throuhg the segmented state space. Moreover, we demonstrate that Reinforcement Learning is accelerated by setting a sub-reward at the important states found by our method in a simple video game learning.
[1] Andrew G. Barto,et al. Automatic Discovery of Subgoals in Reinforcement Learning using Diverse Density , 2001, ICML.
[2] Maja J. Mataric,et al. Reward Functions for Accelerated Learning , 1994, ICML.
[3] J. Grefenstette. Credit Assignment in Rule Discovery Systems Based on Genetic Algorithms , 2005, Machine Learning.