Dragon 3.1 Released at KomodoChess.com

[lkaufman]

Hi all,
I am new here, and I had a few questions for Larry and Mark. How strong/weak is Dragon MCTS compared to regular Dragon, SF and Leela? I had read a few years back that there was an objective to replace the AB Komodo with its MCTS version - is it still on? Also, how useful is Dragon MCTS for CC play? How do we use it correctly for the same? Does the regular hash matter as much as the MCTS hash?
Thanks a lot!

When used just to find the best move, Dragon MCTS is a bit weaker than standard Dragon (just check the rating lists to see the elo difference), though it is still stronger than almost all other engines not derived from Stockfish or Lc0. We hoped (and still hope) that it could surpass normal Dragon, but so far that hasn't happened, they have advanced roughly in tandem. However the MCTS version is stronger when both are used in MP3 or more mode, because displaying the top 3 or more moves costs nothing in MCTS mode but slows down Dragon (or Stockfish or any AB engine) considerably. Also MCTS is more likely to be useful for determining what move will cause human opponents the most practical problems. Since correspondence play is mostly engine play nowadays, that may not help in correspondence play. Using a big MCTS hash is important for long analysis on many threads.

[Werewolf]

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?

[CornfedForever]

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?

+1
I've wondered actually 'how' this can be. I say that because when I am preparing an opening, on move 11, I may see where an engine gives (let me just keep it to 2 moves):

11. Ng5 +.55 but when you look at the responses to that move, you see multiple moves that Black is going to find pretty easy to play, so +.55 is likely and the best I can get.

11. Bb2 + .40 BUT this approach leads Black into more of a quagmire of 'reasonable' choices where it may be really hard to find the thread that leads to +.40...more choices (no not outright blunders or allowing mate) which might have more perfectly reasonable options which are likely to give me a +.95 or +1.10 edge.

As chess is a game of mistakes...I love looking for moves which are not the 'top engine choices'...yet make things more difficult on my opponent.

[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.

[mclane]

A/B is IMO a wrong strategy. It works if the opponent has no plan.
And is a computer that is also doing A/B.
But if the opponent has a plan where he sacs something FOR the plan, A/B has problems to find out. MCTS not.

[Werewolf]

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?

[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.

[Werewolf]

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?

[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.

[Werewolf]

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.