How Ghost AI Adapt and Replicate Complex Mechanics in Fighting Games Using Deep Reinforced Learning
Dinda Lestari Soeprijadi, Anabel Valerie, Ivan Sebastian Edbert, Alvina Aulia · 2024
The Proximal Policy Optimization (PPO) algorithm is used in this research to provide a novel method of teaching artificial intelligence (AI) within the fighting game genre. Our work focuses on creating a Ghost AI that can learn and adjust to intricate mechanics through player behaviour analysis. Even with limited resources, training could only complete 500,000 iterations, or half of the planned one million, but the AI nevertheless showed impressive learning abilities. It created unanticipated tactics that weren't included in the original training set, such a slide attack. This emergent behaviour implies that the AI was successful in winning games after recognizing and taking advantage of patterns in the opponent's gameplay.