Dragon 3.1 Released at KomodoChess.com

[lkaufman]

Also MCTS is more likely to be useful for determining what move will cause human opponents the most practical problems.

I'm sure you've explained this before, but why is this?

The simple explanation is that normal A/B search assumes that the opponent will always play the move the engine thinks is best thruout the search, while MCTS assumes that he will pick somewhat randomly from all reasonable replies. So if some of those reasonable (but inferior per the engine) replies turn out to lose to some deep variation, the engine will get some credit for these possibilities with MCTS, but not with A/B.

Is that very different to Alpha Beta in multi PV? As one works through lines AB multi PV would show alternatives and score them appropriately, how is MCTS different?

Completely unrelated. MultiPV shows alternate lines for the computer at the root, whereas MCTS is all about assuming that moves thruout the tree, including opponent's moves, are not in general totally predictable.

Yes, but as you scroll through lines wouldn’t AB multi PV effectively become similar as you visit each position which interests you?

This interests me, but could you show a practical example from a position where MCTS highlights something AB MPV does not?

Which root move you are looking at is a totally different question from what move many plies down the line the computer will expect the opponent to make. With A/B, the program claims to be able to predict this. With MCTS, it recognizes that it cannot predict the choice when multiple moves appear to be plausible. I don't have time to look for examples, they would be extremely common but difficult to say who is right.

OK, so is this a fair summary: MTCS doesn't assume consistently perfect play from the opponent through all the variations, but only reasonable play. It therefore chooses moves to play for itself which give the opponent more chances (with reasonable, not optimal play) to go astray?

If that's true that is pretty cool. I have to admit I haven't really seen that in my own analysis sessions but I'll look out for it more now.

Yes, that's pretty accurate. It's not that obvious though since the assumed "reasonable" moves by the opponent are usually pretty good ones even if not optimal; that's why Komodo MCTS still outrates almost all other CPU based engines not based on Stockfish or Komodo Dragon.

[magicianofriga]

I've seen Dragon MCTS find sequences of moves that force the opponent to find a narrow sequence of moves to hold a draw. Can this somehow be used in CC play to find ways to make the opponent work for the draw compared to the draw death we are seeing nowadays?

I've also seen a trend where it loves to exchange queens. Why do you think this happens? Does it see that there are potential troubles without queens?

[lkaufman]

I've seen Dragon MCTS find sequences of moves that force the opponent to find a narrow sequence of moves to hold a draw. Can this somehow be used in CC play to find ways to make the opponent work for the draw compared to the draw death we are seeing nowadays?

I've also seen a trend where it loves to exchange queens. Why do you think this happens? Does it see that there are potential troubles without queens?

MCTS should tend to force the opponent to find a narrow sequence of moves to hold a draw whenever possible, so this is clearly helpful against an opponent of lesser strength, particularly an over-the-board human opponent. But if the opponent is also using Komodo Dragon (or Stockfish), then this strategy is unlikely to show benefits. It is possible that MCTS is actually stronger than regular Dragon at say one day per move, but there is no data on this question, only the fact that the elo gap between regular mode and MCTS shrinks as the time limit gets longer.

Regarding exchanging queens, I don't know the reason, perhaps queen trades tend to lead to positions where the moves are less obvious, so more chances for opponent to pick an inferior move. Just a guess.

[Chessqueen]

Stockfish is closing the lead in TCEC ==> https://tcec-chess.com/

[CornfedForever]

It is possible that MCTS is actually stronger than regular Dragon at say one day per move, but there is no data on this question, only the fact that the elo gap between regular mode and MCTS shrinks as the time limit gets longer.

Would this imply the possibility that lengthy analysis of positions (like you might find with Aquarium...or just someone checking on moves in various positions overnight in Chessbase, etc) MCTS might somehow be a bit more beneficial than regular Dragon? Possibly not, but your comment has me wondering.

[lkaufman]

It is possible that MCTS is actually stronger than regular Dragon at say one day per move, but there is no data on this question, only the fact that the elo gap between regular mode and MCTS shrinks as the time limit gets longer.

Would this imply the possibility that lengthy analysis of positions (like you might find with Aquarium...or just someone checking on moves in various positions overnight in Chessbase, etc) MCTS might somehow be a bit more beneficial than regular Dragon? Possibly not, but your comment has me wondering.

It is quite possible. To get the answer, someone would have to run a match against recent Stockfish on good hardware at some very long time control, like game in 24 hours with 10 minute increment for example, using an unbalanced opening book like UHO22 (otherwise 100% draws), once with Dragon MCTS and once with regular Dragon. That would require quite a commitment of hardware time and electricity.

[magicianofriga]

Hi Larry,
Thanks a lot for your response. As I understand, the longer MCTS thinks, the stronger it becomes relative to A/B engines? Is this a trend you have observed in Leela (which is also a MCTS engine)? Does this mean that at VLTC MCTS can surpass Stockfish?

The MCTS option doesn't work with Aquarium IDeA and Nibbler GUI. Is it an issue with the GUI or the engine?

I have a few questions I would like to ask you. Can you please send me your email address so that I can contact you directly (over private chat)?

Thanks a lot for your contributions to chess and I hope that you can spare some time for me!

[lkaufman]

Hi Larry,
Thanks a lot for your response. As I understand, the longer MCTS thinks, the stronger it becomes relative to A/B engines? Is this a trend you have observed in Leela (which is also a MCTS engine)? Does this mean that at VLTC MCTS can surpass Stockfish?

The MCTS option doesn't work with Aquarium IDeA and Nibbler GUI. Is it an issue with the GUI or the engine?

I have a few questions I would like to ask you. Can you please send me your email address so that I can contact you directly (over private chat)?

Thanks a lot for your contributions to chess and I hope that you can spare some time for me!

I am pretty sure that dragon MCTS gains relative to A/B engines with more time, at least in the range for which we have data. I don't think that is the case with Leela, but Leela is very different from Dragon MCTS, because Dragon is a hybrid between A/B and MCTS, whereas Leela is just MCTS. Leela gets a much higher nps count than Dragon MCTS due to using GPU (but much lower than Dragon standard mode), which is probably the reason that it doesn't benefit from longer time controls the way Dragon MCTS does; once you look at a few thousand nodes or so in MCTS, the gains from further doublings don't seem so dramatic. I doubt that Dragon 3.1 MCTS would beat current SF even at VLTC (like hours per move), but we don't have any data on this. I don't know the reason that MCTS option doesn't work with the two GUIs you mention, perhaps they require output that MCTS doesn't supply. Regarding my email, send me a private message here.

[Albert Silver]

Hi Larry,
Thanks a lot for your response. As I understand, the longer MCTS thinks, the stronger it becomes relative to A/B engines? Is this a trend you have observed in Leela (which is also a MCTS engine)? Does this mean that at VLTC MCTS can surpass Stockfish?

The MCTS option doesn't work with Aquarium IDeA and Nibbler GUI. Is it an issue with the GUI or the engine?

I have a few questions I would like to ask you. Can you please send me your email address so that I can contact you directly (over private chat)?

Thanks a lot for your contributions to chess and I hope that you can spare some time for me!

I am pretty sure that dragon MCTS gains relative to A/B engines with more time, at least in the range for which we have data. I don't think that is the case with Leela, but Leela is very different from Dragon MCTS, because Dragon is a hybrid between A/B and MCTS, whereas Leela is just MCTS. Leela gets a much higher nps count than Dragon MCTS due to using GPU.

The GPU is not used for the MCTS search per se, it is used to consult the NN for each position. The actual search is still done on the CPU.

[Chessqueen]

Hi Larry,
Thanks a lot for your response. As I understand, the longer MCTS thinks, the stronger it becomes relative to A/B engines? Is this a trend you have observed in Leela (which is also a MCTS engine)? Does this mean that at VLTC MCTS can surpass Stockfish?

The MCTS option doesn't work with Aquarium IDeA and Nibbler GUI. Is it an issue with the GUI or the engine?

I have a few questions I would like to ask you. Can you please send me your email address so that I can contact you directly (over private chat)?

Thanks a lot for your contributions to chess and I hope that you can spare some time for me!

I am pretty sure that dragon MCTS gains relative to A/B engines with more time, at least in the range for which we have data. I don't think that is the case with Leela, but Leela is very different from Dragon MCTS, because Dragon is a hybrid between A/B and MCTS, whereas Leela is just MCTS. Leela gets a much higher nps count than Dragon MCTS due to using GPU.

The GPU is not used for the MCTS search per se, it is used to consult the NN for each position. The actual search is still done on the CPU.

Whatever they are using in TCEC it is working for Komodo Dragon ==> https://tcec-chess.com/