Help me test my chess bot
The last couple of weeks I've been fooling around with different ideas for how to build a chess bot that's fun for beginners to play against. I don't have much real experience with chess, so I've...
The last couple of weeks I've been fooling around with different ideas for how to build a chess bot that's fun for beginners to play against. I don't have much real experience with chess, so I've mostly just been testing it against myself.
After looking into the different techniques that are used to force engines to play at a certain level, I put together my own (somewhat silly) approach and have had some fun playing against it. The problem is, as I don't really know what I'm doing when it comes to the actual game, I can't be a particularly good judge of how others will feel playing with it.
Regardless of your own skill level, I'd be super appreciative if anyone would give it a try and let me know what they think.
I'm working on a full write up of how it works, but here's the short version:
Click to view the hidden text
The inspiration came from this paper which describes a "Tutoring Search" wherein the engine attempts to find the worst move available that it predicts the opponent will not recognize as an error.
My implementation doesn't follow this exactly, but it does have the same aim. Two engines are used: one (Stockfish) as an oracle treated as a true measure of any state's quality, another (Maia) as a substitute for the opponent model. On each move the bot consults both of these to identify a move which:
- Would plausibly be played by a skilled human, judged by its probability of being played by Maia.
- Provides an advantage to the opponent, judged by Stockfish.
The idea is that, if a human would be likely to play the move, they also would be unlikely to identify it as an error. The ultimate goal is a bot which gives the player plenty of opportunities to win, but only if the mistakes are likely to go unnoticed.
There are a few other supplements to the implementation like adapting to opponent choices and some tweaks to early and end-game play, but the above is the core idea.