Beating Go thanks to the power of randomness
summary
Claims that random simulation plus tree search beats old approaches using Monte Carlo Tree Search, deploys 50k+ playouts per second on strong engines, and wins tournaments historically dominated by heuristics; Crazy Stone beats pros with a 4‑stone handicap, Zen wins one game against Franz‑Josef Dickhut by 1.5 points, and heavy compute (an 80‑core cluster) helps. Papers show 44–51% professional move prediction from a convolutional neural network. A playable Ruby engine Rubykon exists and integrates MCTS; practical enhancements and RAVE nuance whether this actually works in match play.
AI-written summary · about
chapters (17)
- 0:00Start
- 11:51GO vs. Chess
- 12:23Complex vs. Complicated
- 17:04Why is Go so hard?
- 17:20Almost every move is legal
- 18:55Global Impact of moves
- 20:39Evoluation function
- 21:41What is Pi?
- 27:02Exploitation vs Exploration
- 30:19Characteristics
- 30:37Generate a valid random move
- 36:02Enhancements
- 37:31Expert Knowledge
- 41:22Selection
- 44:13What have I learned?
- 44:37Modularizing small components
- 44:55Benchmark everything
related talks
36mPlaying Go with Clojure
18mJavaScript and Artificial Intelligence
40mWhat did AlphaGo do to beat the strongest human Go player?
28mGetting started with open source game playing AIs
25mBuilding Board Games with Ruby
01h 03mIEEE CoG 2019 - Industry Day - David Silver (DeepMind)
46mAI Takeover
33mShall We Play A Game?