Improving reinforcement learning using temporal-difference network EUROCON2009

摘要

Reinforcement learning has been one of popular learning methods for many problems in many different domains. The important point for this method is how fast and efficient it is to learn a new problem. In this paper, we present a new approach to increase the efficiency of the reinforcement learning method with the great help of a predictive model of the problem's environment called temporal-difference network along with observation. This TD network is nourished with the knowledge extracted from another problem with the same task using TD network. First a reinforcement-learning agent tries to learn its environment for the task of wall following. After that we train temporal-difference network (TDN) with intervening observation in the brain of the agent in order to gain a predictive model of the environment. Later the most promising sequences of action-observation of the given environment will be extracted as knowledge to strengthen the reinforcement learning problem in a new environment. Finally this knowledge helps the reinforcement procedure to produce more efficient results.