Incorporating Image Prediction Model in Value Function for Deep Reinforcement Learning

Abstract

Reinforcement learning is an unsupervised learning method that enables an agent to learn its behavior via interaction with the environment. By maximizing the value that represents the expected reward over a certain period of time, the agent can learn to perform the required action. To obtain a high value, selecting the optimal action in an unknown future state is necessary. If an unknown future state can be predicted in advance, better actions can be performed. Therefore, obtaining a high value as a result is possible. In this study, we use a deep learning-based future image generation model to predict unknown future states in advance. By predicting the future state, selecting actions that lead to a higher value is possible. Thus, higher rewards can be expected at an early stage.