ZenD play review

[badukJr]

ZenD, quick play version of Zen, was online the last two days. This was the first time he was on since beginning of December.

ZenD played as a 5d on KGS. His (ranked) record was 72 W - 30 L. An astonishing 102 games in 48 hours.

The scary thing is, is that ZenD has some big flaws to work out still. He is not very good at ko, and his L&D is not great sometimes.

[Mike Novack]

The scary thing is, is that ZenD has some big flaws to work out still. He is not very good at ko, and his L&D is not great sometimes.

You need to not think of these programs as a human playing the game as a human might. The algorithm used isn't analyzing situations locally but only "what move makes it most likely that I win the game" (regardless of what happens here locally).

"He is not very good at ko"
Be more specific. Yes I think the algorithm might have some problems where multiple simultaneous kos are involved (drastic increase in the depth of the tree because some sequences of moves leave the state of the board unchanged).

"L&D is not great sometimes"
What do you mean? If you mean not immediately making a move that lives or not immediately making a move that kills but leaving the situation to be resolved later in exchange for a play elsewhere then that isn't "poor L & D" but disagreement with you about which play is best for the game overall.

Operating one of these programs in a mode where you are shown the moves it was considering (at least the top few) might convince you of this. Not that it didn't see or consider making the key move in this local group but that in the end doing so did not win out as the best move overall for the sake of the game.

I'll repeat, the MCTS based programs are only evaluating the game over all. Not in terms of local situations.

[daniel_the_smith]

Unless they have special code (as they do now for ladders, IIRC), MCTS programs effectively will have half the depth when reading kos. Humans think only about threats at that point, but MCTS programs have to find the move that continues the ko, too. If that were handled, though, I'd actually expect MCTS programs to be great at evaluating threat sizes.

[hyperpape]

This may not be quite what the original poster meant, but it does appear Zen struggles in situations involving ko: http://computer-go.org/pipermail/comput ... 04566.html.

[badukJr]

The scary thing is, is that ZenD has some big flaws to work out still. He is not very good at ko, and his L&D is not great sometimes.

You need to not think of these programs as a human playing the game as a human might. The algorithm used isn't analyzing situations locally but only "what move makes it most likely that I win the game" (regardless of what happens here locally).

"He is not very good at ko"
Be more specific. Yes I think the algorithm might have some problems where multiple simultaneous kos are involved (drastic increase in the depth of the tree because some sequences of moves leave the state of the board unchanged).

As the other posters mentioned, ko reduces the search space, and Zen often plays ko threats before the ko starts, or plays multiple threats in a row before taking the ko back. If this was really optimal then I think we'd see it more often in professional play (It does happen sometimes in pro games, but it happens in almost every single Zen game with a ko)
And as you mentioned, simultaneous kos reduce the search space even further...

"L&D is not great sometimes"
What do you mean? If you mean not immediately making a move that lives or not immediately making a move that kills but leaving the situation to be resolved later in exchange for a play elsewhere then that isn't "poor L & D" but disagreement with you about which play is best for the game overall.

There are many times when Zen plays on a dead group. If the problem is complex, he may not find that the group can be killed in enough searches in the tree, and leave it alone. I feel that a 5d Human will have a greater understanding of L&D, because it is easier for them to cull the search space to a local problem.

Operating one of these programs in a mode where you are shown the moves it was considering (at least the top few) might convince you of this. Not that it didn't see or consider making the key move in this local group but that in the end doing so did not win out as the best move overall for the sake of the game.

I'll repeat, the MCTS based programs are only evaluating the game over all. Not in terms of local situations.

Yeah, I know how they work, and read through some of the papers. How they evaluate the game leads to those weaknesses above. If enough playouts don't find critical points, they will do something weird or wrong.

[Mike Novack]

Yeah, I know how they work, and read through some of the papers. How they evaluate the game leads to those weaknesses above. If enough playouts don't find critical points, they will do something weird or wrong.

I've been waiting a while to see if somebody else wanted to express an opinion before giving what I think the problem is. First of all, I don't think this is just about ko but probably also affects the proper evaluation of moves that break (would break) ladders that do not already exist and probably some other subtle aspects of the game.

Not enough playouts. That's what we have to discuss. If this algorithm is given "enough" time then the sample set on all relevant paths would be sufficient for proper evaluation (note this is different than something like brut force chess evaluation in terms of what unlimited time means). But because there is a time limit only a finite approximation can be used. How well that works depends upon "the state of the game". What I am suggesting here is that for some situations the finite approximation converges toward the ideal more quickly than in other situations.

Premature surrender of ko threats is incorrect evaluation of a move that has a relationship to a ko that does not yet exist (and whose likelihood of materialization depends on the relative balance of ko threats) and the same thing applies to ladders. In these situations the materialization of the ko or the ladder is spread out on many paths and so the sample size along the relevant paths grows slowly. In other words, random addition of paths to the tree means these situations do not quickly converge to the ideal algorithm (unlimited time).

What, if anything, could be done about this I'll leave to those developing the programs.