Re: Objective function to include learning away from draws for more interesting and better chess

[glbchess64]

You must read this : https://lczero.org/blog/2020/04/wdl-head/

And assimilating draw to boring chess is a bit excessive : a lot of people love KGA that generally is a draw with best play. In fact a lot of people consider open games as exciting but according to Leela they give less winning opportunities than closed or semi-open games. In fact most people prefer and find exciting games with fire on the chess board (a lot of exchanges and complex tactics) that are generally easy quick draw for the best engines and think that complex positional play (that gives more winning chance) is boring.

Moreover winning games are also generally merely games where one of the opponent miss something. I personally prefer games where both opponents play the best moves. And these games are often draws.

Le samedi 28 novembre 2020 à 17:30:45 UTC+1, refle...@gmail.com a écrit :

Professional chess' objective function is broken, resulting in many many draws. The space of draws in chess is a massive attractor, and if you reward draws, it is possible to find them. We should instead encourage the engine to learn a space away from draws. You can think of it this way: in exchange for a much smaller set of excellent games which are discovered, we must endure craploads of garbage, boring chess (draws). There is a sweet spot, which I will describe here, that the engines are overlooking by targeting professional tournament-style play. This style of play will also benefit grandmasters.

From the perspective of tuning your engine's objective function, you need to punish draws. This is likely to be problematic for the engine, resulting in catastrophic interference and less or even no learning. To get learning to continue to work, you need to introduce artificial time constraints (constraints on the depth the engine is allowed to analyze, making it much shorter). This will force the engine to gamble on high risk position play that results in sequences of trades which are difficult to analyze, but in which there tends to be a definite winner. 

Interestingly, this objective function will actually benefit from rewarding the engine for losing, at least from time to time, a behavior typically only found in people.

Sincerely,

Brian Mingus

Deep learning / AI / human psychology expert

https://scholar.google.com/citations?user=T_sFnwoAAAAJ&hl=en