Comment by thedstrat | Hacker News Reader

Comment by thedstrat | Hacker News Reader

Parent Full thread

Agreed. The analyses based solely on a comparison of average centipawn loss is so so flawed. It only takes using an engine move once or twice to completely demolish a much better opponent in a game. These types of analyses don't find this type of cheating.

A much better analysis imo would be trying to find the probability of someone at his ELO finding surprising moves. EG I played a 1900 online recently who happened to completely turn around a game by setting up a forced mate in 6, with several branching moves a few moves down which all happened to result in mate because of incredibly lucky piece positions. I can't calculate the probability of someone at a relatively low level like that finding such a move, but I bet it's very low. This is the type of analysis which I'm guessing Magnus is using to assess Niemann as a cheater.

>It only takes using an engine move once or twice to completely demolish a much better opponent in a game.

That's not true. Pick an engine, set to a few hundred points above your strength, then try to beat it using only one or two moves from an engine. You will lose nearly every game, because so many of your other moves will be so below the other player that the 1-2 good moves cannot make it up.

This is demonstrated quite often by the games where GMs are "helped" by others in multi-player games, and it shows that help against a much better opponent takes far more than 1-2 moves.

Among really close players it helps. But not once you get a few 100 ELO points apart.

Regan's analysis works, in part, the way you suggest is better. He looks for tricky moves with non-obvious consequences and looks at one's success rate in those. For the Niemann-Carlsen game, he identified two such moves and Niemann chose suboptimally for both of them. For one is those, Niemann played a nice that could have cost him a tempo. Definitely an inaccuracy.

Niemanns moves fit perfectly for someone using a slightly dated version of Stockfish.

Reply on news.ycombinator.com