Arimaa Forum - Print Page


    
      
        Arimaa Forum
        (http://arimaa.com/arimaa/forum/cgi/YaBB.cgi)
      

        Arimaa >> Bot Development >> Sharp P2 having long thinks?
        
(Message started by: Swynndla on Jun 13^th, 2011, 11:11am)

Title: Sharp P2 having long thinks?
Post by Swynndla on Jun 13^th, 2011, 11:11am

I was looking at this game:
http://arimaa.com/arimaa/gameroom/opengamewin.cgi?client=1&gameid=187321&role=v&side=b where bot_Sharp2010P2 was taking mostly between 10-30 seconds per move, but sometimes it would do 1 minute, 2 minute, and even some 3 minute moves (see for example: 11g, 207g, 298g, 218g). Now "P2" means that it is only looking 2 ply ahead (its move and opponents move) with no extensions, right? So why the long thinks? Unless the server had some really high loads or something (although the long moves had short moves either side). I think there was maybe one other live game going on at the same time (when I was watching towards the end).

I looked at previous games for this bot, and the same thing (up to 3 minute moves occasionally), but other P2 bots I looked at don't do this. So unless my spot-checks have been too few, it looks like it's how this bot is (and not the server load). Perhaps it has some sort of evaluation bug where it gets in some sort of loop, and it's not really thinking beyond 2 ply?

Title: Re: Sharp P2 having long thinks?
Post by lightvector on Jun 13^th, 2011, 11:46am

For Sharp, P2 means that the main full-width search goes 2 turns deep. There's still a quiescence search below that examines capture and goal threats and defenses. A P2 search will go at least 12 steps deep for lines that involve capturing pieces.

Sharp2010 has a problem where the quiescence search can sometimes cause the search to explode, if there are a lot of capture or goal threats (either in the current position, or that would occur after the moves it's considering). This problem also still exists in Sharp2011, although to a lesser degree.

Also, judging from my impressions of other bots, Sharp2010 might be more heavy-weight than many other bots. The evaluation function does a lot of work and is bulky and slow, so the search can't get quite as deep as other bots. So when you force it to get to the same fixed depth (P2), it just takes much longer on average to do so than other bots do.

Title: Re: Sharp P2 having long thinks?
Post by Nazgand on Jun 13^th, 2011, 11:59am

289g, you say? Whoever plays a game that long is three halves as crazy as I. :)
The move that bothered me most for it's long thinking was 207s; I suspect it was looking at eV<^ M>, and analyzing the resultant camel/cat threats. But I don't really know, and 3m5s seems too long.
At 217g, I thought I might actually win, but after thinking for 2m32s it came up with a clever response. I feel that this response time was good for the situation; then again, a better player might have seen it quicker than Sharp.

Edit: Perhaps the reason 207s took as long as it did was the fact that 207g was the first time it was offered a chance to reduce material, and no longer care about previous position history. Perhaps the abrupt reduction of position history is not well managed.
Edit: noticed lightvector's post about extra search for captures.

Title: Re: Sharp P2 having long thinks?
Post by Fritzlein on Jun 13^th, 2011, 12:44pm

on 06/13/11 at 11:11:55, Swynndla wrote:

Now "P2" means that it is only looking 2 ply ahead (its move and opponents move) with no extensions, right?

It depends on the bot. BombP2, and CluelessP2, for example, have extensions beyond two ply, whereas GnoBotP2 and AamiraP2 don't, if I remember correctly.

Title: Re: Sharp P2 having long thinks?
Post by Swynndla on Jun 13^th, 2011, 7:11pm

Ahh thank you all for your responses, I understand now. Thank you lightvector for the explanation of Sharp too.

I was under the impression that P1 meant only one ply with no extensions, and P2 meant only two plies with no extensions. But for a bot that does heavy pruning, that would make for a weak bot, so I see why extensions are turned on. Hmmm, what about a bot that has lots of extensions, eg qsearch, but also extended searches for "interesting" lines, or lines where it feels it needs to search some attack or defense (quiet but powerful) moves (and other extensions maybe also) - wouldn't that sort of bot be just about as strong as its CC version?

And sorry I got the moves wrong above, I meant 10s, 206s, 207s, 217s - but it looks like people knew what I meant. :)

(Edit: Off topic: Nazgand: I think you missed a forced win in two after Sharp's 3 minute think on 220s)

Title: Re: Sharp P2 having long thinks?
Post by lightvector on Jun 13^th, 2011, 8:27pm

A random interesting thought: for versions of bots that are actually quite slow, even if they are configured to be "hardware independent," the bot's actual playing speed might affect human performance in various ways (slow bot might bore humans and make them pay less attention, or give them more time to think, fast bot might unconsciously encourage players to play faster as well, etc.). As the hardware changes the speed of the fixed-performance bots, these factors might change too.

It's probably not that significant an effect, giving the multitude of other confounding factors that are inevitable when playing online games. But it's a thought. ;D

Title: Re: Sharp P2 having long thinks?
Post by Nazgand on Jun 14^th, 2011, 7:00am

Swynndla: do you mean Rf^< E>d>? silver would have responded with c>>>>.
Edit: hmmm. responding with E<R<<^ would be win in 3.

Title: Re: Sharp P2 having long thinks?
Post by lightvector on Jun 14^th, 2011, 11:19am

Nazgand: Pull the dog west instead, looks like goal in 2

Title: Re: Sharp P2 having long thinks?
Post by Nazgand on Jun 14^th, 2011, 3:10pm

No, because rd>> df> would refreeze the rabbit, and silver could use the last step for e>, causing a goal threat. Thus, that is not goal in 2 either, and I find it harder to read.

Edit: I forgot the elephant would freeze the rabbit, disregard my argument.