Hacker News Reader

Comment by opportune | Hacker News Reader

The neural networks are for the evaluation at each node, traversing the game tree is done by using the NN to evaluate how good the position is on the tree’s nodes; how it traverses the tree is minimax with a-b pruning.

In my original comment I didn’t mention pruning or other heuristics like estimating the value of intermediate states because it’s more of an optimization (a very important one!) that doesn’t affect the theory very much.

View on HN

AlphaZero uses Monte Carlo Tree Search not MinMax. So it does use a neural network for the search. Modern Stockfish is the one that moved to having a neural network for position evaluation while keeping MinMax with AlphaBeta and heuristics for the search.

Oops! You are right, I got alphazero and stockfish mixed up.

Nitpick: MCTS is a heuristic search algorithm as well (just like minimax) that doesn't necessarily need a neural network to operate and the MCTS algorithm existed way before AlphaZero, coming from Reinforcement Learning literature. Deepmind's innovation was to replace two of the 4 classic steps of MCTS (selection step and simulation step) with neural network-based policy and evaluation network.

That's a good nitpick. In recent computer chess discussions MCTS is usually shorthand for the search that was novel with AlphaZero and then copied by LeelaChessZero. But you're absolutely right, MCTS doesn't even need to imply a neural network.

Ah, ok, so Minmax/Alpha-Beta to traverse the game tree and NN to evaluate, my mistake (apparently I am the one who should visit the link I provided :)

No, it was I that was too lazy to visit the link, or we both were. I got alphazero and stockfish’s implementations mixed up

Reply on news.ycombinator.com