Show HN: Teaching a Neural Net to Fight (blog.lukesalamone.com)

In an intriguing experiment, a developer has created "Neural Knockout," a neural network trained to play a fighting game using reinforcement learning (RL). This project explores whether AI can develop emergent strategies while playing, revealing both successes and limitations. Initially, the agents struggled with basic game mechanics, such as approaching and attacking opponents. To enhance their learning, the developer employed reward shaping techniques, providing incentives for closing the distance and penalizing damage taken. This led to the agents successfully learning to approach opponents and engage in combat, although exploitative tactics emerged, necessitating the introduction of a grab mechanic to counteract overly defensive behavior. The significance of this project lies in its innovative approach to training RL agents through a league system, promoting diverse experiences rather than relying on self-play. This method proved effective, as the trained model exhibited superior performance against previous iterations that were trained solely via self-play. The implications extend to game development, where the co-evolution of games and AI could help identify game-breaking strategies or bugs prior to human testing. By leveraging RL in this manner, developers can ensure more balanced gameplay, enhancing the overall player experience.