It has been proposed, and we are seriously considering it, to switch to SinglePV mode - i.e. only exploring and showing the single best principal variation according to the engine. Since this ultimately affects you, our visitors, we are leaving the decision in your hands, via the poll below. Here is a brief summary of the advantages and disadvantages of both options:
SinglePV advantages:
- Engines are much stronger in SinglePV mode than in MultiPV mode. This is because they don't waste time exploring suboptimal positions, and also because some powerful optimizations are only enabled in SinglePV mode (e.g. aspiration windows). And when we say "much stronger", we really do mean "much". According to our internal testing, Stockfish 1.9.1 in SinglePV mode is about 230 Elo points stronger than in MultiPV=4 mode.
- For people who are new to computer chess, having more than one variation is confusing. They will often start arguing that the engine is useless because it shows bad moves, which usually sends our blood pressure through the roof :)
- A single analysis line occupies less screen space than four. More room for the chat box, yay!
- Having analysis for more than one move is useful. For example, it's common to ask yourself "why not move X here?", and when X is among the 4 best moves, you get that answer.
- Sometimes you can feel smarter than the engine.
Vote here and help us decide:
6 comments:
Is it possible for there to be two windows, one showing multi PV, another showing single PV and have less chat space. That way, users can close the window that they do not want, and all sides are satisfied (I know it is a little more work for you!)
@cafestream: No, this is not possible. Either one or the other option must be chosen, and will apply for everyone.
Is it an idea to try singe PV for a month or so? I think I prefer multi PV, but I don't know for sure...
Is there a strength difference between
(a) multi-PV mode to a fixed depth d, vs.
(b) single-PV mdoe to the same depth d?
From the testing of a 230 Elo difference with Stockfish in 1-PV mode vs. 4, I infer yes. Going to 4-PV at fixed depth d should take at most 4x the time to do 1PV at depth d. (I actually observe the extra time to be less, a factor more like 3.) Taking 4x the time translates to about -100 rating points, using the conversion I've heard said about Rybka:
() doing 2x as much computation ~= 50 Elo
() adding one more search ply ~= doing about 2.4x as much computation ~= 70 Elo.
Thus multi-PV mode would be depressing the strength of the engine by about 130 Elo. Is this basically right?
Vas Rajlich suggested to me the idea of stepping ahead each legal move and running each in single-PV mode to depth d-1. Ideally this would need pruning to cut off bad moves, and I do not know how to script this easily and record data e.g. in Arena. Is this a workable solution, running this way to some basic depth d achievable in seconds, and then single-PV on the best move found?
The "multi-PV cap" feature of Rybka 3-4 does this pruning automatically (I set it to 4.00), and is the single biggest time-saver (about 3x) and reason I use multi-PV mode and Rybka 3 in my own automated-analysis work.
Fortunately the 4 lines seem to have it 2 to 1 but sadly, most users don't get it. Showing any 'engine' analysis prior to the move being played is absurd...after,it can be interesting to see whether there is a match to the move played...also after an evaluation of the actual move can be interesting particularly if it differs from top choices.
Even more useful would be "what-if-this?(then-this)" human analysis...ideally interactive...with up to a dozen reasonable alternatives analyzed up to 9 ply where necessary, which would be well beyond the horizon of more than 2/3 of users :)
Post a Comment