Is OGS AI good at yose play?

General Go Discussion
1

I intend to study AI review on my yose play and just want to double check. I can’t see a reason AI is not a better yose player than me.

1 Like
2

Depends. The review AI is LeelaZero, which has only one objective—to increase win rate. Increasing score is not an objective, so it can slack sometimes when the win rate would not change. On the other hand, its policy net is plenty strong, and the best move is often among its top choices anyways.

2 Likes
3

That’s good enough for me. Thanks!

4

I believe that yes, Leela is good at yose. But - I don’t think you can learn from it.

All modern Go bots are based on win chances.

If the game is imbalanced by more than a few points, there is no difference between move A, which wins 100% by 5 points, and move B, which wins 100% by 4 points. As a human, you want to learn that A is better because next time it might decide the game.

Bots can not explain their moves.

How are you going to find out why its move is the best? There is certainly a reason. The endgame is very rigid and you can work out a perfect solution if you know how and why. An endgame book will explain this.

Real games do not make good endgame problems.

The difficulty is all over the place and you don’t get feedback.

My suggestion: get a book instead and do some endgame problems.

5 Likes
5

Agreed. Understanding the concept is critical and an endgame book is my next to read. I am thinking of looking at AI’s options and try to make sense of from what I learn from the book. Thanks!

2 Likes
6

Though correct, Golaxy (proprietary) and KataGo (open source) are both modern AI trained with score maximization as a secondary objective and put some effort into finding better-points moves in the endgame. Installing KataGo is a bit tedious and takes some computer savvy, but I got it working on my system.

5 Likes
7

Bots can show you future simulation after every move you make. So you can do experiments on continuations of single move by changing few next moves as you like.
Human reviewers able to tell you very few additional information to that. And its unlikely they will spend their time on all your “tree of moves” experiments.

1 Like