Machine Learning for Boardgame AI in Race for the Galaxy
summary
Race for the Galaxy uses a temporal difference neural network that trains itself by playing ~30,000 games but gains over a million turn-level training events, producing surprisingly strong, non-heuristic play and novel tactics. The system ships as about 12 bifurcations (one per player/expansion config) and two functions per branch — eval board to predict win likelihood and eval role to predict opponents — and runs thousands of forward simulations per turn to pick moves.
AI-written summary · about
chapters (8)
- 0:00Start
- 1:01What Is It That Makes Games Very Replayable
- 5:49Multiplayer in Vr
- 7:26Heuristic Driven Ai
- 16:35Neural Network Structure
- 33:10What Is a Neural Network
- 35:12The Endgame Condition
- 43:13Learning Resources
related talks
48mMachine Takeover - Improving Your Game With Machine Learning & AI
22mDeveloping AI in Racing Games Using Reinforcement Learning
29mThe Present & Future of AI and Games
49mML: Teaching machine to learn from its own mistakes (...)
01h 40mBuilding machine learning pipelines that scale
39mBuilding self evolving video game AI in Python
15mAI Self - Learning Game Playing
31mDesigning game AI with reinforcement learning