NNue, Reinforcement learning • page 1/1 • General Chess Discussion • lichess.org

Where are the details, the documentation. maybe here we could gather the public links focussing on understanding what is behind the title induced questions.

Related questions about the spirit of open source code projects, now that global optimization and training concepts are being implemented (with the borderline confusion between test sets and training sets, but that is not priority).

Is reproducibility a criterion for how open an open source code project is open? How about documentation quality about, e.g. the training method, so that reprocibility can be attained. Can source code about call data sets be enough without the data sets, and how they were generated.

What does it mean here. What kind of reinforcement learning. How was the NN training separated from the autotuning of other parameters. Is the classical eval function being tweaked whilte the NN patch is, do they train on same training data. optimised toward what kind of loss function (can it be written in math? or do we need a call graph simulatoin?).

Main question: how was the training data generated, where is it. how was it used in what kind of reinforcement learning. was it self-play generated? Any thing might help understanding.... by non-developpers perhaps... or should it be another mysterious zone in the chess world...

A lot of these questions apply to lc0 also. They have been more detailed about training. but the training data, although reproducible by huge infrastructure is difficult to host on github.... which asks the general question of data in open source code project where there is a lot of it...

Where are the details, the documentation. maybe here we could gather the public links focussing on understanding what is behind the title induced questions. Related questions about the spirit of open source code projects, now that global optimization and training concepts are being implemented (with the borderline confusion between test sets and training sets, but that is not priority). Is reproducibility a criterion for how open an open source code project is open? How about documentation quality about, e.g. the training method, so that reprocibility can be attained. Can source code about call data sets be enough without the data sets, and how they were generated. What does it mean here. What kind of reinforcement learning. How was the NN training separated from the autotuning of other parameters. Is the classical eval function being tweaked whilte the NN patch is, do they train on same training data. optimised toward what kind of loss function (can it be written in math? or do we need a call graph simulatoin?). Main question: how was the training data generated, where is it. how was it used in what kind of reinforcement learning. was it self-play generated? Any thing might help understanding.... by non-developpers perhaps... or should it be another mysterious zone in the chess world... A lot of these questions apply to lc0 also. They have been more detailed about training. but the training data, although reproducible by huge infrastructure is difficult to host on github.... which asks the general question of data in open source code project where there is a lot of it...

NM MrPushwood

Looks like somebody left one of their index cards laying around.

selfbrain

Yes, there is a difference between open source and open code pasta. Most projects today are the second.

jomega

There is a difference between open source and open collaboration.