HN Explorer

Can someone help me come up with anti-cheat chess AI?

1 points by gilded-lilly ↗ HN
So the idea is fairly simple. Feed an AI millions of bot games and millions of human games, in order for the AI to assign a “human-ness” probability to a given chess move. This is based on the idea that only one or two bot moves would be enough to affect the outcome of a top-level chess match. I assume this would be fairly achievable for someone with a basic idea of how to implement an ML algorithm. Any takers?

2 comments

[ 3.7 ms ] story [ 17.8 ms ] thread
raxxorraxor
Did only a shallow read into the cheating drama, but I guess future matches need support from a proctologist.

I wouldn't know how an AI could determine cheating, you might end up with a probability of a move being supported by an AI. But that is probably not enough.

gilded-lilly
My idea rests on the fact that there are such things as “non-human” moves, I.e moves that a human could not calculate. It’s these that we’re interested in, and an AI should be able to “learn” to detect them, (as these types of moves will occur much less frequently in genuine human games.)

It can be achieved this way - take real games, and pick a random point in the game. Either replace the move with a bot move, or leave the original move. If it is possible to discern a difference between the play styles of humans and computers, even if only in certain situations, then the AI will be able to.

Crucially, this AI would be able to differentiate between good moves that humans and bots agree on, and good moves that only the bots generally make. Of course, you would need to feed the AI mostly 2200+ rated games for this to have any chance of success.