Al-Driven Gameplay: overcoming open space video games with reinforcement learning

This research explores training an AI agent in video games using continuous actions on a continuous space, an area that needs more exploration. Using Unity's ML-Agents package, an agent was trained with a PPO algorithm in a custom video game environment made from scratch, testing different artificial intelligence models programmed in Python. Hyperparameters, rewards, game difficulty and model choices and how they affected the agent's success were studied in different tests. Each of these tests started with a hypothesis that is contrasted with the results of a training session. The conclusions drawn from these results influenced the hypothesis of the subsequent test. Through testing across six versions of increasing complexity, our agent learned to perform better than human players, reaching average rewards of 80 in the final version. Tests with eight players showed that while humans learned game mechanics faster, the trained agent achieved more consistent results.The work shows that accessible AI tools can successfully train agents that match or exceed human performance in complex game environments. The results highlight how model-free algorithms can work effectively in continuous action spaces, offering insights for future game development.