Acceleration of game learning with prediction-based reinforcement learning - toward the emergence of planning behavior

摘要

When humans solve a problem, it is unlikely that they use only the current state of the problem to decide upon an action. It is difficult to explain the human action decision strategy by means of the state to action model, which is the major method used in conventional reinforcement learning (RL). On the contrary, humans appear to predict a future state through the use of past experience and decide upon an action based on that predicted state. In this paper, we propose a prediction-based RL model (PRLmodel). In the PRL model, a state prediction module and an action memory module are added to an actor-critic type RL, and the system predicts and evaluates a future state from a current one based on an expected value table. Then, the system chooses a point of action decision in order to perform the appropriate action. To evaluate the proposed model, we perform a computer simulation using a simple ping pong game.