search

lorserker / ben

a game engine for bridge
GNU General Public License v3.0
41 stars 30 forks source link

Thought on training #56

Closed Muppet1 closed 7 months ago

Muppet1 commented 1 year ago

Not a bug.

Currently we train on a lot of hands which seems like a good idea. One possible problem down the line when we are trying to make better nets is that it will see a lot of opening bids and very quickly these become close to 1. My suggestion is to only train on a few opening bids relatively and once a certain threshold is reached skip them.

when I thought of this a while back I also realised it would work for chess. I jumped onto the stockfish server and oddly someone had thought of it one month before so I may have read that and forgotten I guess. Anyway skipping some opening moves improved even stockfish a bit, and I think the gains in bridge would be more than chess as there are fewer openings.

Obviously I have no clue how to actually do this so I raise it as a thought for the future.

lorserker commented 1 year ago

@Muppet1 thanks! i like the idea

ThorvaldAagaard commented 7 months ago

I have implemented bidding_coverage.py, that limits a given bidding sequence to 10. I am not sure 10 is the right number, but the results seems fine, and the current model is trained with this option.

So I will close this