Deep Reinforcement Learning in Video Games

Abstract

The high interactivity and instant feedback characteristics of games are highly compatible with the trial-and-error learning mechanism of Deep Reinforcement Learning (DRL). In recent years, DRL has achieved a series of landmark breakthroughs in Game AI, from Atari games to superhuman levels in complex multi-agent game environments. With the integration of Large Language Models (LLMs) with DRL, especially Multi-Agent Reinforcement Learning (MARL), DRL applications scenarios are no longer limited to performance score, but have gradually evolved into more complex situations requiring social reasoning and human-machine collaboration. This paper reviews the application of DRL methods in game AI with some famous games, which can be classified into three categories of games according to the type of games, such as classical Arcade games, First- Person 3D games and Multiplayer Competitive games. This paper conducts a detailed review of the evolution of DRL approaches. In addition, there are some challenges existing in this field, like AI agents’ generalization and human-machine collaboration capabilities. This paper also discusses current multi-agent games and some emerging methods that combine LLM with MARL, while outlining future research directions. This review aims to provide researchers with a clear technical overview.