How I trained an AI Model to Beat the 1990's Arcade Game Double Dragon
summary
A distributed training pipeline uses OpenShift + KubeRay to run ~100 emulator instances across 24 A100s, training a PPO-based agent that reliably clears level one and sometimes beats the boss; the model learns a corner-and-elbow tactic. Observation and reward come from memory-scraped pixels and direct Game Boy memory addresses (distance traveled is the primary reward). Code, configs, and saved episodes are stored in S3 for iteration and debugging.
AI-written summary · about
chapters (10)
- 0:00Introduction
- 3:13AI beats Double Dragon
- 7:08OpenShift AI infrastructure
- 11:30Emulator integration
- 17:07Training runs and metrics
- 20:24Neural network setup
- 25:27Action and observation spaces
- 28:17Memory manipulation and rewards
- 30:05Summary and future work
- 32:15Q&A
related talks
35mHow I Built an AI to Beat the 1990s Arcade Game Double Dragon
31mUsing Python to build an AI to play and win SNES StreetFighter II
46mAI Takeover
35mLet's play Space Invaders!
17mTraining My Rival in Java: A Deep Q-Learning AI to Play Azul
56mCan I do AI?
54mTeaching a Computer to Fight Itself (Implementing an OGRE AI)
47mHow we use Reinforcement Learning (...)