The dumbest neural network there is

https://github.com/nescitus/LizardBrain

This engine contains the dumbest neural network for chess evaluation that exists. 784x16x1 (the same size as Tomitank, which was a direct inspiration), with extremely dumb tuner caring only for the input layer. Because of working on a single laptop, I took a shortcut and initialized some weights in a way mimicking piece/square tables.

Current iteration earned a dubious distinction of playing at 2000-2100 level while using a 2800 Elo-capable search.

If it proves capable of learning something more, I'll consider resurrecting Rodent with Tomitank-style neural network adjustement of its eval. If not, view it as an educational toy showing what a real neural network does.

would you consider training a human neural net, since rodent was such a nice sparring trainer, there's the kingbase, and caissabase games, and i have around 60M human games in my db from lichess, chess.com, OTB combined

also its that strong just by figuring out pst through net?

texium wrote:

also its that strong just by figuring out pst through net?

Judging from the games, it learned some pst, some king tropism and blocking enemy passers, all with a minimal 768x16x1 net. This is enough for 2600 Elo, 250 points lower than a handcrafted eval function, so realistically speaking, it is close to what can be achieved with pst alone. It's funny that early iterations prefered Rf1 to 0-0, perhaps I should train the net from scratch starting from king and pawn parameters only.

nescitus wrote:

texium wrote:

also its that strong just by figuring out pst through net?

Judging from the games, it learned some pst, some king tropism and blocking enemy passers, all with a minimal 768x16x1 net. This is enough for 2600 Elo, 250 points lower than a handcrafted eval function, so realistically speaking, it is close to what can be achieved with pst alone. It's funny that early iterations prefered Rf1 to 0-0, perhaps I should train the net from scratch starting from king and pawn parameters only.

just to understand, did you train from nothing, where all it had was 0 eval terms, and only search.

texium wrote:

just to understand, did you train from nothing, where all it had was 0 eval terms, and only search.

I tune mostly from small random values. And I cheat a bit, because input neurons connected with 2 of 16 hidden neurons get initialized with material values and small centralization bonus.

LizardBrain is now equal to Fruit 2.1 and using hybrid evaluation. Mobility, attacks on pieces and king safety are evaluated using handcrafted eval. On the other hand, all the attempts to add pawn evaluation failed so far, which means that neural network got this part right.