Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Does not avoid being mated. #2

Open
hans-ekbrand opened this issue Feb 12, 2019 · 1 comment
Open

Does not avoid being mated. #2

hans-ekbrand opened this issue Feb 12, 2019 · 1 comment

Comments

@hans-ekbrand
Copy link
Owner

hans-ekbrand commented Feb 12, 2019

Does not avoid being mated.

position startpos moves e2e4 e7e5 g1f3 b8c6 f1b5 a7a6 b5a4 g8f6 e1g1 f8e7 f1e1 b7b5 a4b3 d7d6 c2c3 e8g8 h2h3 c8b7 d2d4 f8e8 b1d2 e7f8 a2a4 c6a5 b3c2 e5d4 c3d4 b5b4 b2b3 g7g6 c1b2 f8g7 a1b1 f6d7 h3h4 a8c8 f3g5 g7f6 d2f3 f6g7 e4e5 d7f8 d1d2 b7f3 g5f3 c7c5 d4d5 d6e5 b2e5 g7e5 f3e5 f8d7 e5d7 d8d7 h4h5 d7d6 d2d3 g8f8 h5h6 c8c7 e1e8 f8e8 b1e1 e8d8 c2d1 c7a7 e1e3 a7e7 e3e7 d8e7 d1e2 f7f5 g2g3 e7d7 e2f3 d6b6 d3e3 b6d6 f3g2 d6e7 e3d3 e7f6 g2f3 f6b6 g3g4 b6f6 g1f1 f6a1 f1g2 a1f6 f3d1 f5g4 d3g3 f6d6 d1g4 d7c7 g3f3 c5c4 f3f7 c7b6 f7h7 c4c3 h7g8 c3c2 h6h7 c2c1q h7h8q a5b3 g8b8 d6b8 h8b8 b6a5 b8d8 a5a4 d5d6 c1f4 g4d1 f4d4 d8e8 a4a3 d6d7 d4d1 d7d8q d1g4 g2h2 b3d4 e8e3 d4b3 d8b6 a6a5 b6e6 g4d1 h2g3 a5a4 e6e5

In the position black is worse but can still defend for many moves, yet the moved played allowed white to instantly mate.

This might not be a problem with terminal nodes, it might be related to the cpuct part of the global weights, with that turned on, I have seen it make a totally random move, which perhaps can happen when some bad move gets just enough policy boost to get a instant increase in global weight. If that happens just before it is time to move, then things the move selection get almost random. Perhaps we need some guard against that kind of instability (or just turn down the cpuct coeff a bit).

@hans-ekbrand
Copy link
Owner Author

hans-ekbrand commented May 3, 2019

I saw a new instance of this with rev 7f9e20e, which is very recent: 29 april, so the problem is not solved, and NN 32603 (a classic NN, I use for testing all the time).

position startpos moves d2d4 g8f6 c2c4 c7c5 d4d5 e7e6 b1c3 e6d5 c4d5 d7d6 e2e4 g7g6 g1f3 f8g7 f1e2 e8g8 e1g1 f8e8 f3d2 b8a6 f2f3 a6c7 a2a4 b7b6 g1h1 d8e7 d2c4 c8a6 c1g5 h7h6 g5h4 e7d7 c4e3 a6e2 d1e2 a7a6 e3c4 a8b8 h4g3 g7f8 e2c2 b6b5 c4a5 f6h5 g3f2 f7f5 a5c6 b8a8 b2b4 f5e4 f3e4 c5b4 c6b4 f8g7 a1e1 g8h7 a4a5 a8c8 f2g1 d7g4 c2d1 g4d1 c3d1 g6g5 g2g3 h7g6 e1e2 c7a8 b4a6 c8c1 a6b4 c1c4 b4c6 a8c7 f1e1 h5f6 g1b6 c7a6 c6d4 f6d7 d4b5 g7e5 d1e3 c4b4 b5a3 e5c3 e1c1 c3b2 e2b2 b4b2 a3c4 b2e2 c4d6 e8b8 d6c4 a6c5 b6c5 d7c5 d5d6 c5d3 c1a1 d3f2 h1g1 f2h3 g1f1 e2h2 f1e1 h3f2 e1e2 f2e4 e2d3 e4c5 d3c3 b8d8 c3b4 c5e4 c4e5 g6g7 d6d7 h2b2 b4c4 g7f6 e5d3 b2b7 e3d5 f6f5 d5b6 h6h5 a1f1 f5g4 c4d5 e4d2 d3e5 g4g3 f1g1 g3f4 d5e6 d2f3 g1f1 g5g4 e5c6 b7d7 b6d7 d8e8 e6d6 h5h4 d7f6 g4g3 f6e8 g3g2 f1a1 h4h3 a5a6 h3h2 a6a7 h2h1q a7a8q g2g1q a1g1 h1g1 a8a4 f4g3 e8f6 g1e3 a4g4 g3h2 g4c4 h2g2 f6d5 e3g5 d6c7 g2g3 c6e7 g3f2 c4c5 f2g3 c5c2 g5e5 c7d7 f3h4 c2d3 g3h2 d5e3 e5f4 d3e2 h2h3 e2e1 f4e4 d7d6 e4f3 d6c5 h4f5 e7f5

In this positition glow will prefer f3e2 and ponder e1h4 and claim q ~ 0. However e1h4 is mate, so it is as if she thinks mate is draw.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

1 participant