What did AlphaGo do to beat the strongest human Go player?
summary
AlphaGo beats Lee Sedol 4–1 by combining Monte Carlo tree search with deep learning: a 12‑layer convolutional neural network trained on human games (≈55% move prediction) plus a fast rollout policy and a self‑play reinforcement learning policy that refines play. It uses a value network to estimate win probability, mixes rollouts and value estimates during selection, runs on many GPUs/TPUs, and produces surprising results from a huge search space (branching factor ~250, state space ~10^171).
AI-written summary · about
chapters (4)
related talks
01h 03mIEEE CoG 2019 - Industry Day - David Silver (DeepMind)
37mA Funny Thing Happened On The Way To Reimplementing AlphaGo
46mBeating Go thanks to the power of randomness
21mWhen the machines out-smarts us
56mCan I do AI?
19mKeynote: AlphaStar: How Google Infrastructure Made it Possible
40mWhere Machine Intelligence Ends and Human Creativity Begins
14mAI for Science