Combining Imitation + Reinforcement Learning to win the Bot Bowl competition
summary
A hybrid agent called MimicBot combines behavioral cloning from a scripted Fantasy Football AI with reinforcement learning updates to stabilize training and avoid policy collapse, claiming the Bot Bowl III win. It uses a specifically designed deep policy network, leverages imitation pretraining on scripted-play traces, then alternates RL updates to raise win rate. The system handles 20,000 input scalars and 8,000 actions, runs in about 20 ms on a laptop CPU, and reports much higher sample efficiency than pure RL.
AI-written summary · about
chapters (9)
- 0:00Imitation vs reinforcement learning
- 1:10Why combine them
- 3:59Imitation learning deep dive
- 7:15Reinforcement learning challenges
- 13:12Deep RL explained
- 17:06Bot Bowl difficulty
- 20:59How Mimi Bot won
- 25:49Key takeaways
- 26:30Q&A on real-world reinforcement learning
related talks
48mPractical Model-based Algorithms for Reinforcement Learning and Imitation Learning, with...
01h 17mTraining intelligent game agents using deep reinforcement learning
37mA practical perspective on training reinforcement learning bots.
31mDesigning game AI with reinforcement learning
14mHuman-in-the-Loop Reinforcement Learning
27mUsing Reinforcement Learning to Optimize IAP Offer Recommendations in Mobile Games
31mFrom Single AI to Multi AI Systems
32mThe Road to General AI