Hacker News Reader

Comment by throw7 | Hacker News Reader

You either have evidence or you don't. This quasi economic modeling of chess players is ridiculous. He says his models don't know anything about playing chess. OK, let's apply your models to all sorts of areas where we suspect fraudulent play.

This chess drama is corrosive. Speculation from day 1 starting with carlsen.

View on HN

The model described here seems. Like bunk. It set of alarm bells for me when they said it wasn't based on an understanding of chess.

But regarding whether it's possible to make such an argument, I'd refer you to the more compelling case the moderators of a Minecraft speedrunning forum made against the steamer Dream. https://youtu.be/-MYw9LcLCb4

The allegation is that the drop rates of key items were manipulated. It's a statistical argument, but it's founded in the mechanics of the game.

After offering counterarguments and denying it for months, Dream eventually admitted to it, though they claim it was an accident (in a manner which is plausible, though I personally doubt it).

This is a fair point, but I think there is confusion around the meaning of the fact that Regan's models do not "know anything about chess". I believe this refers to prior domain knowledge embedded in the algorithm itself - but the idea of exploiting data from a large number of GM games is precisely so the model can extract domain "knowledge" from statistical indicators. As for "applying the model to all sorts of areas", yes conceptually if your model can infer anything from any kind of data distribution this applies. In practice it just won't.

If you take a step back and look at the AlphaGo project, the story shows us a very similar model architecture could be applied to a different game, chess, with AlphaZero. And the "zero" in this case means the model is build with zero human knowledge - not even GM games, only self-play. These models do entirely different things - A0'model given a position, tries to find the best lines, while KR's tool, given a set of games from a given player, will try to price the probability of foul play - but in both cases, no programmer-introduced chess heuristics are part of the model. Yet AlphaZero performs the way it does and appears to "know" quite a bit about the game. So if a purely statistical approach can yield a very strong engine, which has been qualified as displaying some sort of "creativity" (for lack of a better term) in its play, I don't think it is too far-fetched to imagine other models extracting knowledge about what human play should look like.

Totally agree that this is wrong from Magnus, and cheating from Niemann or not he is responsible for the toxic drama.

>You either have evidence or you don't. This quasi economic modeling of chess players is ridiculous

Modeling provides the best evidence in all aspects of life (design, science, invention, marketing, medicine, ....). All human knowledge is built in modeling and statistical evidence. Nothing is 100% certain except mathematics, and even that is often fuzzier than 100%.

So how is modeling yet another thing ridiculous when it provides empirically the best methods in so many other domains?

Reply on news.ycombinator.com