Dominion AI – Temple Gates

The boardgame app currently in beta, Dominion, represents the first commercial implementation of the techniques behind Alpha Zero, the project outlined by DeepMind’s whitepaper. Dominion combines reinforcement learning and a Monte Carlo tree search with a neural network that weights the branches of the tree search for the first time in a shipping product. The Puzzle Most players play digital boardgames against an AI. It’s a way to learn a new game in a safe, unjudging environment. It’s a way to investigate the space of the design at their own pace. And if done right, an AI can challenge a player, no matter their skill level. The best AIs excel through learning, but practical applications of reinforcement learning for gaming AI have been illusive. While theoretically an AI can play against itself to learn and grow, there have been little to no examples of this working successfully in production, other than our previous title, Race for the Galaxy. The research project, AlphaStar was crea

The boardgame app currently in beta, Dominion, represents the first commercial implementation of the techniques behind Alpha Zero, the project outlined by DeepMind’s whitepaper . Dominion combines reinforcement learning and a Monte Carlo tree search with a neural network that weights the branches of the tree search for the first time in a shipping product. The Puzzle Most players play digital boardgames against an AI. It’s a way to learn a new game in a safe, unjudging environment. It’s a way to investigate the space of the design at their own pace. And if done right, an AI can challenge

related reading

• Simple Alpha Zeroweb.stanford.edu
• Why Tool AIs Want to Be Agent AIs · Gwern.netgwern.net
• Deep Q-Networks Explained — LessWronglesswrong.com
• Frontier Coding Agents Can Now Implement an AlphaZero Self-Play Machine Learning Pipeline For Connect Four That Performs Comparably to an External Solver — LessWronglesswrong.com
• AlphaGo Zero: Minimal Policy Improvement, Expectation Propagation and other Connectionsinference.vc
• OthelloGPT learned a bag of heuristics — LessWronglesswrong.com
• Andrej Karpathy — AGI is still a decade awaydwarkesh.com
• EfficientZero: How It Works — LessWronglesswrong.com
• Building machines that learn and think like people | Behavioral and Brain Sciences | Cambridge Corecambridge.org
• Learning To Play Settlers of Catan With Deep RLsettlers-rl.github.io
• How DeepMind's Generally Capable Agents Were Trained — LessWronglesswrong.com
• Generally capable agents emerge from open-ended play — Google DeepMinddeepmind.google