Designing game AI with reinforcement learning
summary
An engineer describes building a reinforcement-learning agent to play a physical shuffleboard-style game, training it to beat a human without hand-coded strategy. The talk explains how the agent learns using three cooperating neural networks, walks through three successive RL approaches that each fixed previous problems, demonstrates a self-play failure mode where the agent converges to a single move, and includes real code, live demos and practical tips for developers starting RL game projects.
AI-written summary · about
chapters (14)
- 0:00Introduction
- 1:04Game AI overview
- 3:33Reinforcement learning basics
- 5:19Game: Palet breton
- 8:06Actor-Critic architecture
- 9:50Implementation and tests
- 12:20Training optimization
- 14:30Reward design
- 16:34Training vs opponents
- 20:50Balancing strategies
- 23:16Model details
- 26:06Technical challenges
- 27:07Demo results
- 28:51Summary
related talks
24mDesigning a Video Game AI with Reinforcement Learning
56mCan I do AI?
30mReinforcement Learning - ChatGPT, Playing Games & More
01h 17mTraining intelligent game agents using deep reinforcement learning
49mML: Teaching machine to learn from its own mistakes (...)
26mHacking Reinforcement Learning
01h 03mIEEE CoG 2019 - Industry Day - David Silver (DeepMind)
32mThe Road to General AI