Archived talk:User:AaronR 20071111

From RoboWiki
Jump to: navigation, search
       Archive        This is an archive of past discussions. Do not edit the contents of this page. If you wish to start a new discussion or revive an old one, please do so on the current talk page.     

It seems that your Ex-current best robot still is your best bot. Nice eh, that 'logical improvements' don't always pay off? Good luck with your next improvements! -- GrubbmGait

My solution was to run Che's pattern matcher on myself and run as far from those bullets as possible :). (See AntiPatternMatcherChallenge2K6) -- Simonton

WaveSurfing, at best, looks like a really good RandomMovement to a PatternMatching gun. Since Che is using RandomMovement, you're basically on an even playing field with it in the movement department. So I'd say Che's excellent PatternMatcher is probably still stronger than your guns. Best of luck! Also, don't rush to defeat Che too quickly - then you'll have to move on to its big brother, Chalk, which is a million times worse! =) -- Voidious

Thanks for the tip! You are right, my guns are not very good compared to Che's pattern matcher. I'll be releasing that new bot within a couple days, here is what it looks like at present:

  • Completely pluggable framework.
  • Wavesurfing movement started out copied from the wave surfing tutorial bot but now has rolling averages and will move away from the enemy intelligently.
  • Virtual guns array is unsegmented and has a HOT gun, a CircularTargeting gun, and a segmented GuessFactorTargeting gun with rolling averages.
  • Supports melee but is next to useless against more than three enemies - it is really only supposed to fight 1-on-1.

Any suggestions on how to improve it, particularly vs. pattern matchers (yes, it gets creamed by Tron)? Thanks in advance. --AaronR

First, I think all versions of Tron use a kinda unique gun, called TronsGun or DynamicClustering. To be more clear, it was once completely unique, but some of us have built our own versions of it since then (Chalk being the most notable, in my opinion). It's not exactly PatternMatching, so improving your movement against Tron will likely do one thing only: improve your movement against Tron. =)

As for tuning your WaveSurfing to work well against PatternMatching, that's a mystery none of us have truly cracked. It seems the only way to really kick a PM gun's butt is to do AntiPatternMatching. Take a look at the MovementChallenge2K6/Results to see how well some of us do, Che is one of the reference bots. The only really useful tip I think I have for you right now is to consider ditching the HeadOnTargeting and CircularTargeting guns in your VirtualGuns array - you may well perform better with just the GuessFactor gun. (I know very few bots where the GF wouldn't be the best to choose, but your VG array might think the others are better sometimes and use them, thus decreasing your performance.) I'd say at least give it a try and see how it works.

Sounds like you are moving in the right direction, though - modular code, starting with the most important pieces and building onto them. Best of luck.

-- Voidious

I am considering ditching the HOT gun, since GF targeting will do the same thing by firing at guess factor 0. The original version of the bot used only GF targeting without the other two guns but it really performed badly against anything except wavesurfers, and that only because I kept the rolling depth low. Adding the circular gun made a big improvement in the performance. At one point I tried having two GF guns in my array, one with a rolling depth of 2 to handle adaptive movement and another with a rolling depth of 20 to handle everything else, but most of the time they just ended up firing at the exact same bin.

When originally designed the bot, I did the same thing with the gun as I did with the movement: copy the code from the GuessFactorTargeting tutorial, adapt it to my framework, and add rolling averages. Do you have any quick fixes to improve a really simple GF gun? --AaronR

EDIT: I just tried changing it back to the 2 GF gun system and it works this time!

Horizon 0.8.6 is now released, and ranks ~94 in the roborumble. Thank you for your help! --AaronR



Alas, it looks like the new version of Horizon does slightly worse overall in the rumble than the old version. See Horizon/Questions for why I think this is the case - in short, the new movement can dodge HOT and advanced targeting, but nothing in-between. Note to self: GrubbmGait says that logical improvements don't always pay off.

-- AaronR

Congrats on the progress with Horizon! -- Simonton

There are no threads on this page yet.
Personal tools