TalkChess.com

leavenfish wrote: ↑Fri May 17, 2019 8:06 pm

lkaufman wrote: ↑Tue May 14, 2019 7:08 pm

leavenfish wrote: ↑Tue May 14, 2019 5:52 pm

lkaufman wrote: ↑Sun May 12, 2019 8:23 pm

konsolas wrote: ↑Sun May 12, 2019 6:23 pm

lkaufman wrote: ↑Sat May 11, 2019 11:48 pm ~snip~

The scores reported by an MCTS engine like Komodo MCTS already reflect the possiblility of the opponent going wrong in plausible ways. Maybe it doesn't do it the same way you would, but at least to a significant degree it does what you want.

That's really interesting. If this is possible with MCTS, have you considered adding a mode to Komodo which adds more weight to the possibility of the opponent going wrong, which would allow people to find trappy lines in analysis?

If regular Komodo did that, it would just become a much weaker cousin of Komodo MCTS. That's the fundamental difference between standard ("Alpha-Beta") engines and MCTS engines; standard ones assume that the opponent will always play the move the engine considers best, whereas MCTS assumes that all reasonable moves have some chance of being chosen. Komodo MCTS (and Lc0 and spinoffs if you have suitable GPU) are the engines you should use.

There seems to be no real info on the Komodo website about vs 13.

Is there some info you can point me to about exactly how MCTS really works? I would like to see how, as you say, it 'already reflects the possibility of the opponent going wrong in plausible ways'. If it in some ways already does what I do manually for opening anaylsis, then it might be worth getting.

Otherwise, if you just look at certain tests that assign ratings for results, there has been little to no real improvement since K11.3.1 (perhaps even regression according to some) which I already have. The only thing 'new' being MCTS - which at least in 'play' is even weaker, but I am only interested in an analysis partner.

For analysis, there are two reasons you might want to use Komodo 13.01 MCTS. !. If you analyze in MultiPV Mode, Komodo MCTS is stronger than normal Komodo, because being forced to analyze multiple lines to get a score cripples all Alpha-Beta engines but is free for MCTS. 2. MCTS assumes that there is some probability that every legal move will be played at any point, with the probability depending on how good the move appears to be. So if other replies are inferior but not "obviously bad" that will make your candidate move look better. The more plausible ways there are to go wrong for the opponent, the better your move will look.

Thanks - is there anywhere noted the changes from 12 to 13? I presume the usual 'speed-ups' and eval tweaks, but I cannot even find that listed anywhere. I did find on the Komodo website: "Komodo 13 MCTS is much better than previous versions in its ability to utilize hardware with many cores. " but that would seem largely irrelevant to the average consumer with your average hardware.

The changes from Komodo 12 to 13 in MCTS mode (which is where most of the changes were made) are so vast and pervasive that it is hopeless to even try to enumerate them. The elo gains are in the 300 to 400 range on any number of threads, so it's really a very different engine. Virtually every aspect was improved, some of the changes being rather radical. Most cannot be explained without describing in detail how Komodo MCTS works, which is pretty unique. As for regular Komodo, there were quite a few changes, but only a few that show up as measurable elo gains on one to four threads. To be honest, if you never use MultiPV or MCTS mode, don't have more than a few cores, and already own Komodo 12, I wouldn't recommend spending the money for Komodo 13 unless the cost is insignificant to you. But I think that most people do use MultiPV, and if so Komodo 13 MCTS is clearly better than Komodo 12 in either mode.

Thank you Mark and Larry,

The progress you guys are making with the MCTS version is interesting to monitor.

Couple of days ago i made up this position. And i find it quite amazing that Komodo is capable of holding this position as white, whereas SF loses as white. Komodo know white got long term compensation, even black is pure exchange ahead. The defensive outpost on d4 and in addition of opposite-colored bishops make this position a headache for stockfish. There is no entrance path on the queenside for the rooks. According a friend of me (GM rated 2500) he suggested black's plan is to play on the kingside and said the position is probably lost for white.

[d]r1bqk2r/pp3ppp/8/3p4/3Np3/1P2P3/PB1P1PPP/3Q1RK1 b kq - 0 1

Gabor Szots wrote: ↑Thu May 09, 2019 9:58 am

Ovyron wrote: ↑Wed May 08, 2019 10:48 pm

Gabor Szots wrote: ↑Tue May 07, 2019 8:32 amWhere do I set MultiPV under the Shredder GUI?

It seems that you can't set MutiPV under the Shredder GUI (it hides the setting in engine options, and doesn't allow to set it directly; I had to make my own compile of McCain that calls it "MultiPW" just to fix this.)

The best you can do is using the F6 Key to increase MutiPV lines by one, and the F5 Key to decrease them by one, but it's impossible to play MultiPV games in the Shredder GUI (as it defaults to 1 in play mode.)

Thank you for your reply. I edited my Komodo eng file to include a line MultiPV=6, only I don't know whether it has an effect.

Haha! Whoa, yes, it does! And it works for any engine! This is a clever solution, and I was having to edit engine code and make my own compile to achieve what adding this line on option does

So here's the thing, Shredder GUI doesn't know this is Multi-PV output, so it treats it like normal output. If you're using an older version of Shredder Classic, right-click the analysis panel and tick the "Continuous Display" option so you don't mix anything.

The difference is that on normal output the best move is always shown at the very bottom of the analysis, while with a setting like MultiPV=6, the 6th best move is shown at the very bottom, with the 5th best in the previous line, and the 4th best in the previous line and so on.

I believe this was the reason engines like McCain's Tactical Setting or ShashChess's Deep Analysis Mode just switch MultiPV like this (no different than setting MultiPV in engine options), except they hide the scores of non-best moves so the score of the best one still appears at the bottom (even though for analysis knowing the scores of all those moves for free is very useful, not just finding the best move unattended.)

Thanks for this concept, now all engines that support MultiPV are able to use McCain's Tactical Setting by setting it in engine options

How many rollouts per second does one core of K13.01 achieve? I know this will differ greatly based on processor and processor speed, but is there a rough way to know?

schack wrote: ↑Tue May 21, 2019 10:26 pm How many rollouts per second does one core of K13.01 achieve? I know this will differ greatly based on processor and processor speed, but is there a rough way to know?

\

On my fast laptop, using one core only, it reports 1000 nodes after 12 seconds in opening position, so 83 nps. It doesn't sound like enough to even play chess, but each of these nodes is based on thousands of regular nodes searched, so if we counted nodes that way it might be in the ballpark of a million or so nps on one core.

i need this engine

@larry - Can you explain what the search tree looks like in layman's terms? I'm curious as to how this operates.

schack wrote: ↑Wed May 22, 2019 4:15 pm @larry - Can you explain what the search tree looks like in layman's terms? I'm curious as to how this operates.

With standard alpha-beta search, the programs just explore alone one line at a time, only saving/remembering the best line. With MCTS search, the whole game tree is stored in memory as a linked list of nodes. A node is a position. So the root node is simply the position on the board being searched. Its child nodes are the positions made after making a move from the root. The game tree is expanded a by formula with two parts. One part is simple the average win probability for that position and is the Exploit part of the formula. The other part is an Explore part which helps ensure the program looks at positions with a lower win probability, just in case a deeper search of them leads to something better. You select the best mode and repeat, deeper and deeper in the tree. When you hit a leaf node (with no children) you expand them and return a win probability up the line in the tree you traversed, addin in the win probability of the new node, adjusted for the side to move.

Each node has a bunch of numbers like a count how often this nodes has been selected, The sum of the win probabilities, a number called policy to help select the nodes, perhaps node locking stuff so only one processor works on updating that node at a time and so on.

Mark