LIVE: Reinforcement Learning AI plays Art of Fighting 龍虎之拳 | NULLMICGO

NULLMICGO live: a reinforcement learning AI learning Sega Genesis Art of Fighting (龍虎之拳) from scratch. The fighter being trained is Ryo Sakazaki. The target is 1,000,000 steps. This run just started, so messy play and missed specials are expected. What you are looking at: • Game window: Ryo vs the CPU, live • Terminal: reward, score, Spirit, and move tags on every step • Small window, top right: percent complete, time left, and fps • TensorBoard: rollout/ep_len_mean is how long each fight lasts, not training progress. Progress is the percent in the small window How the model is trained: The agent sees pixels, presses buttons, and is scored on the in-game score plus a custom reward. The algorithm is PPO from Stable-Baselines3. The emulator is stable-retro, running on a Mac with Apple Silicon. This is reinforcement learning, not a language model playing the game. At about 75% Spirit it throws Haoh Shoukou Ken or Ryuko Ranbu. Up close it favors low kicks, it does not stand idle, and at low health it tries to keep distance. Only one game window is open so the fight stays watchable. That makes the picture slow, a few frames per second. The time left in the small window changes as the speed changes. #ArtOfFighting #龍虎之拳 #ReinforcementLearning #AITraining #NULLMICGO #RyoSakazaki #PPO #Livestream