Please play our final class project bot

For our machine learning final project, we trained a chess engine bot to play like a human. Our training consisted of showing millions of positions and the moves played on those positions from past Lichess games (around 1500 rating).

We need your help establishing an accurate ELO!, right now it is rated 1200, but it's primarily played with other bots, and we're not sure we can trust their ratings. So, when you feel like taking a break, please consider challenging our Lichess bot here: @NeuralKnight1500.

We'd love to hear if you have feedback about your experience playing NeuralKnight1500. Feel free to include your ELO (if you have any) with your post below.

The bot will be active today and the rest of the weekend. If you have any issues playing, please let us know below so we can restart our local server.

I do not like the bot >;(

The bot would not checkmate me

It's a lot stronger than 1200, but misses somewhat obvious things occasionally.

The bot plays really strong. I (DWZ about 1800, lichess > 2000) lost two games, then I won. Endgame seems to be its weakness. This is very human-like.
Question: Could your Bot act as a coach and tell people what is good and what is bad? King Safety, Mobility, Pawn structure and so on. Strategic things.
I modified SF 15.1 to give classical values out, but since SF 16 there are only two NNUE. Classic is nuked.
I am a streamer at twitch (PrettyHardy). If you are interested, you could explain your ideas via discord, could be interesting for my community.

Thank you all for your feedback. We're super excited to see how it's performing. Clearly, it has some issues with stalemates and threefold repetitions that we will have to work on.

@ehenkes w.r.t. your question about coaching: our bot was trained to "emulate" a 1500 Lichess player. We succeed insofar as it has the same type of understanding (and it makes the same kind of mistakes) as a 1500 player. If we did a good job, then this should be a good tool for lower-rated players to practice against an engine that makes similar types of mistakes.

To make our AI "explainable" would require a somewhat different type of training, but it would be a natural continuation of our work. It would have to involve some kind of "diffing" between what our 1500-engine wants and what a strong engine (stockfish) would do. Maybe @mszylko has some ideas for how it could work?

Alright, so I just played a game with your bot for your sake and wouldn't mind sharing my thoughts - Undoubtedly, you did a good job emulating the average 1200 rated player as I believe your bot sucks colossal ass.
I applaud you for putting forth effort on this complex project. You do deserve credit for it.

Machine learning is such an interesting field in the world of programming so to speak. Very promising. I assume this is how the creators of Alpha Zero did things before it rose to prominence (I'm not so sure). I believe your project could potentially shoot through the ceiling in the future, after you've made your bot competitive.

Here's the game I played by the way.



Best of luck on your ongoing project

I played a game (FIDE Elo 1900ish). lichess.org/45qXaFXd

Tried a sharp opening, thinking it would crush me there. It kind of did. And then simply got trivially mated.

It was blundering multiple times.

When playing, its thinking time (or lack thereof) was a bit annoying. It indicates a missing search, which explains the tactical unawareness. Does it play human-like moves? Hard to judge in that opening. :-)

I like the concept of the bot, but I think it would be nice to let the bot "think" every move for 2-3 seconds, so it feels more human-like. Here's my game against it: lichess.org/oOVy4TSa/black#15

The bot got confused a lil bit by too many rooks