Comparative Study of Curiosity Learning Methods in Reinforcement Learning for Rogue-like Games

Abstract

In recent years, there have been numerous attempts to optimize an agent's behavior in complex environments such as video games using deep reinforcement learning. However, a common problem in reinforcement learning is that learning becomes difficult when rewards are sparsely given by the environment. To address this issue, a method called curiosity-based learning, which uses intrinsic rewards based on the novelty of observed states in addition to extrinsic rewards, has been proposed. In this paper, we focus on a rogue-like game that has the characteristics of sparse rewards and random environment generation. We will compare the efficiency of Q-learning and Deep Q-Networks(DQN), and with using curiosity-based methods. We evaluate the performance of each method by exploring randomly generated dungeons using the trained models. Then conduct a discussion on the results.