Creating a novel adaptive AI for a Real-Time Fighting game

Abstract

AI has transformed how we play, bringing both online and offline gaming to excit ing new levels.While progress has been made in many areas, research in game AI hasn’t kept up like other fields. In this research, we want to create adaptable AI for video games that learns in real time from both the environment and opponents. Our pipeline consists of two major modules: a Monte Carlo Tree Search (MCTS) module for decision-making and a machine learning module to handle the action list. We use reinforcement learning, specifically Q-learning, to dynamically change the action list, which boosts the AI’s performance. Our results show that this hybrid method not only improves AI speed and efficacy, but also helps it to react in real time, mimicking human strategic behavior. Experiments on limited-resources show that our approach works, with high win rates against a variety of AI opponents. Our approach provides an insight into combining MCTS and reinforcement learning for real-time adaptive AI in gaming situations.