Understanding AlphaZero at Home
Policy, value, search, and the compromises required on personal hardware.
Why AlphaZero?
I love chess, play regularly on Lichess, and wanted to reproduce the central idea behind AlphaZero on hardware available to one person.
Traditional chess engines rely heavily on carefully engineered search and evaluation. AlphaZero combined a policy-and-value neural network with Monte Carlo Tree Search, then improved through self-play.
The policy and value outputs
The policy estimates which moves deserve attention. The value estimates the eventual outcome from the current position. Search uses both: the policy guides exploration, while the value replaces many expensive random rollouts.
The practical limitation
The published system used extraordinary compute. A home implementation must reduce network size, batch inference carefully, store replay data efficiently, and accept that iteration speed matters more than reproducing the original scale.