Lila: Make the AI opponent much weaker in the first levels

[ornicar]

Level 1 should give anyone who knows the rules a chance to win.

Maybe a second issue, but we could look into making the strength relative to the time control.

[greg-finley]

There's an incredible video about Elo rating all sorts of crazy strategies. At here he mentions a strategy where x% of the time you make a move with Stockfish and the rest of the time you randomly pick from the legal moves. Maybe that fits here?

[ddugovic]

It appears there are two simple-sounding things we can try:

1. Reduce the thinking time per move: https://github.com/niklasf/fishnet/blob/29771d5860300c6565971495bb0dbc0c02cf7dba/fishnet.py#L121
2. Try the new UCI_Elo parameter? It works the same way as Skill Level but can be more finely tuned... https://github.com/official-stockfish/Stockfish

[Eugenio-Bruno]

Just a random idea: there are very, very small leela nets, and I think you can set them to do a single evaluation per position. There is a chance it might be low elo but kind of "human". I could set up a lichess bot to try the concept.

[ianfab]

What we did for pychess was to extend Skill Level to negative values in (Fairy-)Stockfish and add a lvl0 in fishnet (as we wanted to keep the existing levels consistent with lichess) that makes use of the negative Skill Level.

@ddugovic Reducing thinking time or switching to UCI_Elo will not help for low levels, because the lower levels are more limited by depth than thinking time and UCI_Elo has the same lower bound as Skill Level.

[ddugovic]

Thanks, such an extension makes sense to me as well:
Extend Skill Level and UCI_Elo lower bounds

I agree the thinking time restriction has far less impact than reasonable thinking time restrictions would have. Although I'd like to encourage trying my latest master, for the past week there have been communication errors between https://github.com/dolegi/lichess-bot and Stockfish with "bestmove" being lost, and I've been unable to discern which program(s) are at fault.

[niklasf]

I deployed latest fishnet master to move servers, which inlcudes @ddugovic's patch for lower Elo bounds. And indeed the lower levels feel much weaker now.

While the parameters could be tweaked, I think this is very close to the best solution that we're going to get without changing engines.

[ornicar]

https://lichess.org/8rg8EJbB

Oh my.

[ornicar]

lvl 2 https://lichess.org/izntppiG

[Eugenio-Bruno]

This is the first iteration in my experiments in making a weak chess engine but a bit more human-like. It's probably lower than 800 right now. It's based on a neural network.

https://lichess.org/EAxJjRWW/

you can also take a look at the last few days of games, as it trained. They're against a wide range of ratings.