@schlawg said in #6:
> I agree 100% that many would love to see it, but we would not add something we felt was fundamentally dishonest.
Exactly this. Also, the repository that you point out classifies good/brilliant moves in a very simplistic manner as others have also pointed out. What's worse than lacking a feature? It is having a version of that feature implemented poorly.
A better, but more complex approach could be something like this. Classify moves as brilliant if they are dubious or consist of a sacrifice, the average centipawn loss is not significant when compared with the best move, and the lines and variations that it follow have a good chance of confusing the opponent to make mistakes. It'd be hard to compute this and will be complex, but even then it will have its limitations. A simple backward defensive knight move can be brilliant too from a human perspective, but it won't be able to detect it.
If you can classify tons of games with annotated brilliant moves of all sorts, then perhaps you could use a ML model to train it with. You can give trusted users the ability to rate the brilliant moves, and then use that as an input to better classify future brilliant moves. I am not sure how feasible or accurate this will be, but seems like it won't be worth the effort. Nevertheless, an interesting task.
> I agree 100% that many would love to see it, but we would not add something we felt was fundamentally dishonest.
Exactly this. Also, the repository that you point out classifies good/brilliant moves in a very simplistic manner as others have also pointed out. What's worse than lacking a feature? It is having a version of that feature implemented poorly.
A better, but more complex approach could be something like this. Classify moves as brilliant if they are dubious or consist of a sacrifice, the average centipawn loss is not significant when compared with the best move, and the lines and variations that it follow have a good chance of confusing the opponent to make mistakes. It'd be hard to compute this and will be complex, but even then it will have its limitations. A simple backward defensive knight move can be brilliant too from a human perspective, but it won't be able to detect it.
If you can classify tons of games with annotated brilliant moves of all sorts, then perhaps you could use a ML model to train it with. You can give trusted users the ability to rate the brilliant moves, and then use that as an input to better classify future brilliant moves. I am not sure how feasible or accurate this will be, but seems like it won't be worth the effort. Nevertheless, an interesting task.