Tabula Rasa Learning Approach Proposal

[severeduck]

@tryingsomestuff have you considered QueensGambit/CrazyAra#212?

[tryingsomestuff]

Minic is a alpha-beta engine using a neural network evaluation function (and not an engine using a network to predict the next move itself). In this sense, the CrazyAra learning process (or LC0, or Maïa, or ...) cannot be achieved in Minic.

But to talk about "starting learning from scratch", Connor, Seer author, did something like this for the evaluation function, starting to learn on how to win with few pieces on the board and then add more and more pieces in the learning process. This is a technic sometimes refered as "curriculum-learning" if you are interesting to read about this.

Another related concept may be about how we optimize the search (using SPSA or other technics like bayesian optimization for instance). Here, starting from an initial good guess or even with dummy values, we try to find the best search parameters combinaison. And to do this we have framework that plays a lot of games and make parameters evolves in what is considered a good direction.

So I guess my answer is both yes and no !