AI that mimics specific player's style

You can download all of their pgn files, and then analyze every single position with stockfish and compute the changes for every move they make. Now feed that data into an ai, and have stockfish analyze some current position. Now, for every move possible, analyze that position, and compute the changes. Now have the ai pick the move such that it thinks the entire game is most likely played by the person whose pgn all you analyzed. It is a 1 or 2 day long experiment it should take to implement all of that, but I don't know how successful it will be in mimicking the pgn. How good it is really hinges on the projections you analyze of all of their games, and how good it is at predicting whether or not a particular game was played by your opponent - I think those are the two variables that matter for an ai mimicking a specific player's style.

Sorry, couldn't really address your topic (I'm still spitting).

@cauffybeen said in #1:

AI that mimics specific player's style, Does it exist?

I don't know how many unique games you'd need to train a chess AI (it might be more than a person would ever play), but, I know that there are engines which allow you to tune them to mimic a player, like Rodent: https://github.com/nescitus/rodent-iv

With these engines, you can select an Elo and weigh things like how often a player uses an opening, or how much they do or don't value a solid position, or a sharp position, or the two bishops, or certain kinds of pawn structures, or certain kinds of imbalances.

So, if you know how a player plays the game, you can get the engine to play a similar sort of game.

Something tells me that choolingo has been reading a lot of Tom Robbins.

@MrPushwood said in #3:

Sorry, couldn't really address your topic (I'm still spitting).

Now that is a loyal citizen of Lichessington

We don't need an AI about this.

Example:: let's say you play D35 QGD Exchange 100 times. Machine learning with a neural network can analyze your tendencies by position in specific variations and decide your most likely move blunderabilitude and all. 🫠

There's @maia1, @maia5, and @maia9, trained on games of 1100s, 1500s, and 1900s respectively. One could probably train Maia on games of a specific player to mimic that player.

See also @EricBot.

Assuming it was possible to make an engine play like a person, what would be the point? There's no shortage of human players. And even if you play against this engine and win, you know you didn't really win because that is just a dumbed down engine, programmed to let you win. When you play against a person you know they're doing their best and when they blunder that's genuine, not just some algorithm that decided it should blunder.