| |
BGonline.org Forums
eXtreme Gammon - Review/Overview
Posted By: Stick
Date: Monday, 29 June 2009, at 11:17 a.m.
This is the second time I'm having to type this out because my FF crashed most of the way through the last posting. Very annoying, so if I cut something short blame the browser.
Last night Xavier Dufaure de Citres, the developer of eXtreme Gammon, was nice enough to meet a few of us online and answer any outstanding questions we had, take suggestions for improvements, etc...
Currently eXtreme Gammon which is $49.95 to purchase comes with only one license. This means you can only use it on one computer unless you pay for multiple licenses. You can however open as many instances of eXtreme as you want on the computer it is installed and licensed on. Xavier said he was rethinking the one license per purchase/computer situation and I'm certain he'll find a solution for those of us who own/use a lot of computers for backgammon work, perhaps a monetary increment per license or a bulk rate for unlimited licenses.
As of now eXtreme also does not do rollouts in parallel as GNU does. This is a top priority in my opinion. For those who may not be familiar eXtremem rolls out one play however many trials, say 1296, then it rolls out the next candidate play 1296 trials, and on down the line for as many candidate plays as there are. (Snowie does the same) GNU rolls the plays out simultaneously meaning it rolls out Play A one trial, Play B one trial, etc.. then Play A the second trial, then Play B the second trial, and updates the results so you're able to stop the results when the joint standard deviation (JSD) [or confidence interval, CI] reaches a statistically significant point. By doing rollouts in parallel the end user is capable of saving a lot of time by stopping the rollouts 'early'.
eXtreme uses an ELO scale to tell you how well or poorly you're playing. There is no equivalent Snowie error rate. Xavier said it's easy to include this and likely will but prefers his method. (of not counting forced moves like GNU or even near forced moves where all plays are of the same equity) I prefer this method also but since Snowie is the de facto ER amongst bg players for better or worse, I think it should be included.
Right now you can finish a straight race at the click of a button. It must be a race though. We suggested changing this to any point in the game if you want to go to the end of it to have that option. Xavier said he will eventually implement it.
You're able to copy and paste a position from GNU into eXtreme by copying the GNUBg ID and pasting it, ctrl + v, onto the eXtreme screen. The inverse is not possible. eXtreme does not output a GNUBg ID to swap positions to GNU. I don't know if this change will be made. You are able to swap positions between different instances of XG using the copy and paste method.
Some other things of note:
Wondering how strong XG is I played it a fair amount of money games on its highest analysis level. This level is referred to as 'XG Roller'. XG Roller performs a 6 ply truncated rollout to choose its moves. I believe 3600 trials but I could be mistaken, I'm working from memory. The delay is minimal, it seems for the most part to be playing just as fast as Snowie or GNU on their higher plies. After getting a feel that XG plays well I decided to allow it to play itself, log the games, and run them through GNU's 2 ply analysis. I allowed it to play hundreds of games against itself, XG Roller v. XG Roller, logged the games, and its error rate according to GNU is impressive enough to assume for money its analysis on XG Roller is as strong as any of the bots. The worst 'session' it had an error rate of 1.17 according to GNU. This, however, was a short session, only 10 games, and all of the longer sessions (>>25 games) XG played under a 1 ER according to GNU.
I have yet to tackle any of the match play stuff not having the time and Zorba had started posting matches of XG also.
As for rollouts, you can't do rollouts using the level of XG Roller. The highest setting you can use for rollouts is 3 ply. There are two different 3 ply settings, one labeled simply '3 ply', the other has the labeling of '3 ply red'. It's clear from how they're staggered that regular 3 ply is stronger than 3 ply red. Xavier said the difference is basically 20 ELO and I could tell 3 ply red's rollouts were noticeably faster in the few tests I've conducted but I feel it's too much to sacrifice and will only run rollouts with regular 3 ply.
I'm still in the testing phase of how fast the rollouts are compared to its GNU equivalents but I have a lot of material to work with. Depending on the position though I think it's safe to say XG is at least 4x faster and up to 10 or 15x as fast. For example, I took Keene's most recently posted cube decision and put it in to both bots for money. http://www.bgonline.org/forums/webbbs_config.pl?read=42554 I toggled both to 324 trials. XG finished its 3 ply rollout in just under 3 minutes (2m52s) whereas GNU had finished 108 trials after ~15 minutes and then I stopped it. That's a ridiculous increase in speed so if accuracy holds that alone would make it worth the price tag in my opinion. I have the feeling that early game positions won't roll out 15x as fast as in my example but I still think they'll be ~5x as fast, depending on the position. I will continue my investigation as much as possible but with the Michigan (Novi) ABT rapidly approaching, then Europe for 4 months after that, I doubt I'll get to do a whole lot more.
I'd highly recommend for anyone serious about bg and using the bots to at least give the trial version a whirl. Simborg swears by its ease of use but honestly it didn't seem any tougher or easier than the other bots to use. There are some whistles (not bells so much) that may make it a more fun experience for your casual user and I won't go into those right now. There are also options that come as the default one should change and I think Xavier is working on a FAQ and other various documents for the program.
Anything I may have forgotten I again place the blame on Firefox. Any questions fire away, Xavier said he would check in and a couple of us should be familiar enough with it by now to answer any questions. (MCG, Phil, Zorba) There are of course a lot of minor improvements we could probably think of but for now I'm trying to stick to the bigger points and not overwhelm the developer who put a lot of work into making a speed machine that might be able to churn out just as accurate rollouts and save us a load of time. ASM rocks.
Stick
| |
BGonline.org Forums is maintained by Stick with WebBBS 5.12.