Graph Structure Exploration for Reinforcement Learning State Embedding - Train Tetris Agent with Graph Neural Network.

Artificial intelligence has been developed in many fields due to the rapid spread of the Internet worldwide and the rapid development of computer computing power in the last decade. During this development, many branches of AI, such as computer vision, natural language processing, graph deep learning, and reinforcement learning, have been proposed and studied, making AI ubiquitous in people's lives today. In recent years, cross-domain applications have been a very hot topic, where the combination of graph deep learning and reinforcement learning has achieved good results in several fields. In our previous work, we used graph deep learning as an encoder to encode game states on Tetris game and fed it into a reinforcement learning algorithm for training game agents with success. In this paper, we build on our previous work and explore the impact of different graph construction methods on the performance of the game Agent. We compare the performance between graph structures with different number of edges and finally find that there are cases where the Game Agent can maintain good performance even when using graph structures with significantly fewer edges.