Comment by dmurray | Hacker News Reader

Comment by dmurray | Hacker News Reader

Parent Full thread

> To simplify a bit, engines (whether neural network or classical engines) always play with the implicit assumption that they play against the best response of an equally strong opponent.

This is a bit of a simplification! Neural net engines generally use Monte Carlo Tree Search over a pure minimax/alpha-beta search. MCTS rewards moves that lead to wins in all explored variations, not just the best play for each player.

Your point is still true though: computers still aren't able to reliably find "this will be the most testing line in a practical game against a human opponent". They're too far ahead of us: the MCTS leads to better results against computers of similar strength, but doesn't accurately predict where humans will go wrong. Humans (with computer help) are still better at that.

I considered clarifying that, but it's a bit subtle.

First of all, neural networks are typically trained from (millions of) self-play games. So you have the "equally strong opponent" assumption right there in the weights of the network.

Second, when actually playing, the variations that are explored with high probability in MCTS are (recursively) informed by the same neural networks that are applied at the root. This means that there's no extra reward for moves that lead to complications in lines that only a "stupid human" would play.

In effect, neither alpha-beta search nor Monte-Carlo tree search have any concept of "tricking" a weaker opponent, admittedly for very different reasons and more probalistically in the MCTS case.

That's what I meant with "the implicit assumption that they play against the best response of an equally strong opponent" -- yes it's a simplification, but directionally true I think.

What's the difference ? if you win against best play you win against less good play too no ?

If you win, yes. But perhaps you can find two moves that both draw against best play (overwhelmingly likely in most chess positions you will prepare, though even computers aren't good enough to tell you this for certain). The important thing then is to pick the one that wins against less good play, or perhaps that you won't lose when you inevitably show some less good play.

Reply on news.ycombinator.com