Difference between revisions of "Talk:Pattern Matcher Challenge"

From Robowiki
Jump to navigation Jump to search
(→‎Sequential Prediction: heap instead of sort)
(→‎Sequential Prediction: faster now, thanks)
 
(One intermediate revision by one other user not shown)
Line 17: Line 17:
  
 
: On another note, I currently have a better scoring version (over 102%) but I don't know if it is cheating. I started tweaking how to find pattern bot faster, so I turn to the best angle, but I also turn the gun and the bot so it turns faster. I'm not sure if turning the bot is allowed, I'm currently testing a version that doesn't turn the bot and see if is enough to get the first place. If not I will only post the results as information. And as soon as I have a fully functional bot in the rumble, I will release the side version tweaked towards this challenge as well. --[[User:Zyx|zyx]] 05:35, 16 October 2009 (UTC)
 
: On another note, I currently have a better scoring version (over 102%) but I don't know if it is cheating. I started tweaking how to find pattern bot faster, so I turn to the best angle, but I also turn the gun and the bot so it turns faster. I'm not sure if turning the bot is allowed, I'm currently testing a version that doesn't turn the bot and see if is enough to get the first place. If not I will only post the results as information. And as soon as I have a fully functional bot in the rumble, I will release the side version tweaked towards this challenge as well. --[[User:Zyx|zyx]] 05:35, 16 October 2009 (UTC)
 +
 +
: Yeah, nice job, guys! Always cool to see some new tech at work. Not sure about the bot turn thing - I think "your bot may not move" means it's not allowed, but I never thought about turning. Of course you could just note it in the comment field. --[[User:Voidious|Voidious]] 13:42, 16 October 2009 (UTC)
 +
 +
:: Thanks for the heap tip.  I've eliminated almost all skipped turns by keeping a sorted list of length N;  it uses an insertion sort when adding new items (slow) but discards matches that are shorter than one at the list's tail end, so it's far, far faster than what I was doing before (I really need to learn more about trees one of these days...).  Also I use a very low N when the gun isn't cool yet.  It's worth mentioning that seq-pred is very sensitive to missed scans:  one tick of missing data can really screw up your match lengths. --[[User:Darkcanuck|Darkcanuck]] 22:34, 17 October 2009 (UTC)

Latest revision as of 23:34, 17 October 2009

Archived Talk:
/Archived Talk 20091015

Sequential Prediction

I saw that there was a lot of discussion over at Sequential Prediction and thought I'd migrate this page to see some results compared to some of the originals. Also wanted to show off the score of my new Pattern Matcher... :-) --KID 01:54, 16 October 2009 (UTC)

I'd like to try, but can't get a hold of PatternBot's jar... My simple implementation of seq-pred scores about 65% vs Shadow but that's all I've tested to date. --Darkcanuck 02:21, 16 October 2009 (UTC)
Nevermind, I got it -- repository must have been offline, I'll put a copy on my server to save others this frustration! --Darkcanuck 03:02, 16 October 2009 (UTC)
I've been planning to migrate this page(s) for the same reason, but never got the time to do it. Thanks. --zyx 04:35, 16 October 2009 (UTC)

Those are some really nice results you've both posted! My implementation isn't so good, it starts off well but then ends up skipping too many turns -- probably from sorting the ever-increasing number of matches in the histogram. --Darkcanuck 05:05, 16 October 2009 (UTC)

I have a heap that stores the N best matches, 31 at most, then I keep the best ceil(N / 3) that stay inside the field and weight them by match length. Hope it helps somehow.
On another note, I currently have a better scoring version (over 102%) but I don't know if it is cheating. I started tweaking how to find pattern bot faster, so I turn to the best angle, but I also turn the gun and the bot so it turns faster. I'm not sure if turning the bot is allowed, I'm currently testing a version that doesn't turn the bot and see if is enough to get the first place. If not I will only post the results as information. And as soon as I have a fully functional bot in the rumble, I will release the side version tweaked towards this challenge as well. --zyx 05:35, 16 October 2009 (UTC)
Yeah, nice job, guys! Always cool to see some new tech at work. Not sure about the bot turn thing - I think "your bot may not move" means it's not allowed, but I never thought about turning. Of course you could just note it in the comment field. --Voidious 13:42, 16 October 2009 (UTC)
Thanks for the heap tip. I've eliminated almost all skipped turns by keeping a sorted list of length N; it uses an insertion sort when adding new items (slow) but discards matches that are shorter than one at the list's tail end, so it's far, far faster than what I was doing before (I really need to learn more about trees one of these days...). Also I use a very low N when the gun isn't cool yet. It's worth mentioning that seq-pred is very sensitive to missed scans: one tick of missing data can really screw up your match lengths. --Darkcanuck 22:34, 17 October 2009 (UTC)