[Home]TargetingChallenge2K7/ResultsChat

Robo Home | TargetingChallenge2K7 | Changes | Preferences | AllPages

/ReferenceBots - /HowTo - /FastLearning - /Results - /ResultsFastLearning - /ResultsChat - /PreChat


Somebody should run LifelongObsession & post the results; I don't have a computer anymore. The source is in the jar. Just open up LifelongObsession.java, comment out the getMovement() method, and make the getNextFirepower?() method return 3. Anybody? -- Simonton

GrubbmGait: congradulations on first place!! ;) -- Simonton

GrubbmGait, did you just run the battles against WeeksOnEnd manually? If not, what version of Robocode are you using with RoboLeague? -- Voidious

Oh, and I'll gladly run LifelongObsession sometime. Don't have a computer?! =) That's no way to live! Just kidding... -- Voidious

Oh ... uhh ... I just realized that LifelongObsession doesn't have any memory checks in place ... so ... 500 rounds may not work. You can try like -Xmx2048M or something, but otherwise I guess only /FastLearning will possible. -- Simonton

@Voidious: Thanks for running those bots, now I don't feel like a complete loser anymore ;-) -- GrubbmGait

Caffiene b1.1.2 is a rather advanced version, other then being unsegmented, but nothing that would make it an oddity in actual aiming. I think its a good bottom reference. Perhaps someone can do Che? --Chase-san

I thought it would be so much higher Simonton! Can you send me your challenge bot, i'll run the other targeting challenges (it should pwn the original). --Chase-san

Well, credit to Simonton for having sweet and innovative new PatternMatching guns, but I don't think I'm alone in believing that PM is inherently weaker than other stats-based approaches. Of course, CodeSize is a huge factor among non-MegaBots. It's quite surprising how very close it is to Che, actually, overall and divided among surfers / non-surfers. I'll be curious to see the other TC's, too. -- Voidious

It's one of the first PM's to make the 2000 club, (though i'm not sure but maybe Shaakious, made it first? But looking at the version history, doesn't seem likely). Not only make it, but to stay there instead of crossing the line alot. --Chase-san

This was a new single tick gun. An experiment in using advancing/lateral velocity instead of heading/velocity. Unsegemented. Every little tweak really throws all the scores around. It usually seems to average out, though. I ran some scores for the 2K6 challenge running 2 of them in a VG array & managed to match SeaSerpent's fast learning score (this was a while ago, maybe I can find & post them ... I think there's some others I can post somewhere, too). I'm still not convinced there isn't more potential for pattern matching - I just haven't figured it out yet. But yes, they seem to do very well in CodeSize leagues. -- Simonton

Jeez, how is it we have no fast learning results yet? Running Dookious now, then I'll move onto some of the inactive authors' bots. -- Voidious

Would somebody mind running Decado (FastLearning? and/or otherwise)? Just comment out the setAhead and change bulletPower to always be 3...I'm interested in what it would get. Sorry I haven't managed to get Roboleague running yet, I've been busy with adding segmentation to Druss, and life in general. -- Skilgannon

No problem. Seems reasonably fast in my test run here, so I should be able to get both (regular and fast learning) posted today. -- Voidious

Thanks -- Skilgannon

Hmmn...and I got to 3 in the micro with that...heh. My movement must be good. And the fact that I set my 'best' distance to 400 :-). Of course, I only keep the last 20 000 'frames', so I expect the fast learning result to be better.

 -- Skilgannon -- who will be working on a better gun

I'll try to get the TCCalc script updated with an option for TC2K7 formatting soon. I like the split format we have now, but it's definitely a pain to re-arrange / calculate that stuff yourself. Chase-san, didn't mean anything personal saying it's a pain, just fyi. =) -- Voidious

Now I see the split resultst and I thank God om my bare knees that there are still a lot more non-surfers around. Now I know where my effort should go whenever I get that frustrating GoToSurfing? working like I want it to. I will try to run the FL challenge asap (read mondayevening). -- GrubbmGait

Interesting...Decado has the best score against Shadow in the fast learning :-) -- Skilgannon

True, and thats very cool. Though thats only one bot, and... Shadows gun probably rips Decado's movement to shreads. --Chase-san

Yeah, Shadow does =). It's just surprising considering it gets the lowest of the scores overall. It would be nice to have divisions - ie. Micro, Mini, Mega, at least that way I won't feel like the last cow's tail ;-)-- Skilgannon

Its my ANN map system crudely glued ontop of a gutted Raiko (no movement modified firing power), currently the stats do not decay, which is what I attribute to the horrible surfer scores. It does pretty well against most the non surfers, except the multi-modes at this point, I have some ideas on how to improve it, including adding greater dimensional diversity to my map (its lattice is only currently 3d, I made it infinitely expandable). It only records data on real bullets. --Chase-san

Prototype seems to have a problem against Fortune, GrubbmGrb and WeeklongObsession. It completely obliterates most the surfers, so I need to find a way to fix it to work against MultiMode robots, but I got nothing on that yet, only a few theories. So if you have a suggestion, it would be appreciatied. I can get around to hitting the random mover later. --Chase-san

I enjoyed those 87 minutes of glory while DCResearch had the best /FastLearning score against Shadow ... -- Simonton

Awesome! Nice work, David! =) Very impressive. -- Voidious

Thanks. I actually didn't make any changes to the gun for 0.9, but apparently the 0.860 gun code works well against the reference bots? =P --David Alves

I think Dooki's gun is at as strong or stronger than Phoenix's, but maybe Phoenix is better at hitting these 10 particular bots. --David Alves

MostCompetitiveRobocoderEver, you just want it to be better so you have an excuse to improve your gun some more. ;) :D :P --Chase-san

"Phoenix DC test" - hmm ... I'm beginning to understand your title. Was the thought process something like, "I see people competing; I'd better go win!"? :P -- Simonton

It's a gun I wrote a year ago but gave up on. Since DC guns are all the rage these days I figured I'd dust it off and see what it can do. It's actually not that bad against the non-surfers but it's pretty poor against surfers. --David Alves

I would like to see what scores Engineer gets in this. Just for reference. ;) --Chase-san

ABC - I hope you don't mind, I fixed your math. All I changed was the grand total, based on the two subtotals. But then, maybe they were wrong? Feel free to check out RoboResearch; it's really handy for running challenges like this (and spitting out wiki output to copy-paste). -- Simonton

I used TCCalc and did the subtotals by hand, I'll check out RoboResearch when I get the time. BTW, your bots (and RMX) are the ones keeping our scores down when compared to the top 2... -- ABC

Just as a point of discussion, I am wondering how people feel that the targeting and movement challenges have affected the development of their bots, and what impact it has had on their Roborumble performance. Alternatively, are the challenges just another way to showcase your hard work? -- Martin

One thing I think is very true is that improvement in the targeting challenges is much more likely to correlate to RoboRumble points than improvement in the movement challenges. I think that MovementChallenge2K7 might have been a big step in the right direction, though. I have had many cases where improving my TC scores resulted in RoboRumble rating points, and I generally expect it to, but I've also had a few where it resulted in nothing.

I definitely don't think the challenges are sufficient as your only benchmark when improving your bots, but I do think they are good benchmarks that can really help if used "correctly" - like knowing that improving your TC score against surfers will not help your overall rating much, or that improving your movement against Shadow might only help against a handful of bots. I must say that I think the "showcase" aspect has become a really important aspect of the challenges, at least for me. Really, there is nothing to say that RoboRumble performance is the "true" test of a bot (or a gun or a movement) as opposed to carefully constructed benchmarks, IMO, or even a different style of tournament / ranking (like PremierLeague). Interesting topic.

-- Voidious

It's a very imprecise science, bot improvement. Especially when you're fighting against small details in your gun/movement. Big changes are always easier to evaluate. Personally I've always focused at improving my score against top bots, that's probably why I've always tended to be better ranked in the PremierLeague than in the general rumble. I like the challenges a lot though, and I believe they are very useful for testing changes and new ideas. A big improvement in one of the challenges is always guaranteed to have a positive impact in your rumble rating, smaller ones are not. -- ABC

Perhaps a challenge that is 500 seasons of 1 round matches, no data saving, to get first round preformance, which can be very important. --Chase-san

Nice scores for Gaff, Darkcanuck! 74 against Shadow is crazy - that's 5+ better than second best?? Wow. -- Voidious

Indeed very nice stuff in Gaff! Yeah, random movers ban be of a pain to hit as well as the top guns can, I also have more trouble there than with the surfers. I'd also agree that targeting is certainly easier to test, of course targeting an movement are rather interrelated (and increasingly so at close range I suspect). LunarTwins interestingly performs better with it's simple rolling-average-turn-rate-circular-targeting gun than if I throw a well proven PM gun like Waylander's gun on it. In any case yes, movement might be good for you to work on though, considering it appears to be hurting Gaff's score far more than the targeting. Personally I tend to bounce between targeting, movement, and teams, changing after one category starts to frustrate me too much, but of course, everyone has their own habits. One thing I noticed is that you noted in Gaff's page that you're no longer using hidden nodes. I find this interesting for a couple reasons: Firstly this indicates my theory that Gaff was taking advantage of more complex relationships between inputs isn't actually the case (to a significant degree anyways). Secondly, I'm thinking data dumps of the weights would be far more comprehensible now that hidden nodes are gone (as the weights now far more directly indicate the strength that an input category is associated with an output category). :) -- Rednaxela

Attribute Weight Normalizing method
Lateral Velocity 10 divide by 8, absolute value
Lateral Acceleration 10 if(abs(accel) > 0.4), signum(accel)/2 + 0.5, else 0.5
Advancing Velocity 2 divide by 16 + 0.5
Distance 5 divide by 1200
Distance from position 10 ticks ago 3 divide by 8*10=80
Forward Wall 5 max of 1.5 radians
Reverse Wall 2 max of 1 radian
Time since lateral direction change 3 max of 1.5 BFT, divide by 1.5
Time since decel more than 0.4 3 max of 1.5 BFT, divide by 1.5
Current GF being visited 3 GF/2 + 0.5

Feel free to mix, match and adapt =) But if you figure out any improvements, let us know! Of course, what works for NN won't necessarily work for DC. BTW, DrussGT is ONLY tuned to hit RM, but I think enough of the attributes are wacky enough that surfers have trouble dodging them. -- Skilgannon

Attribute "Weight" Calculation
Lateral Velocity All 4 buffers Absolute value
Bullet Time All 4 buffers Distance / Bullet speed
Acceleration All 4 buffers slowing / same speed within 0.5 / speeding up
Forward Wall Distance 2 (high) buffers GuessFactor to wall
Reverse Wall Distance 1 (highest) buffer GuessFactor to wall (reverse)
Time since velocity change 2 (high) buffers Divided by bullet time
Distance last 8 ticks 2 (high) buffers now.distance(eight-ticks-ago)

Again, this is not NN, but in my case, VisitCountStats; nevertheless, I'd expect these attributes to work well in any statistical system. The AntiSurfer gun doesn't really use any attributes not listed here, just less attributes, less segments, and super fast RollingAverage. The wall distance translates basically to: "at what (non-precise) GuessFactor the opponent will be when he hits the wall, if he simply orbits the wave." Good luck!

-- Voidious

Attribute Range
Distance 0 - 1100
Lateral Velocity 0 - 8 (absolute)
Advancing Velocity -8 - +8
Lateral Accel. -2 - +2
Advancing Accel. -2 - +2
Time since velocity change 0 - 32
Time since CCW/CW change 0 - 32
Wall distance (along heading) 0 - 500

The nice thing about NN is that the net seems to figure out the best segmentation (I was segmenting the data in earlier versions, but performance is actually better using the raw data) as long as the data is scaled from -1 to +1. I'll give the distance-since input a try, that might help against RM.

Skilgannon, what do you mean by the "current" GF -- wouldn't that always be 0? -- Darkcanuck

No, it's the GF of the wave that is currently hitting them, ie. the hit GF of the wave that is being removed at the time that this one is being fired. And just to be sure, your lateral accel is based off of the absolute value, right? -- Skilgannon

Interesting -- firing wave only, or any wave? (Gaff learns from both, with a heavy emphasis on firing waves) Yes, the lateral accel. is corrected to match the absolute lateral velocity. That's the bug I had originally, wasn't taking the absolute nor correcting the accel., have no idea why 1.10 performed as well as it did. Probably the hidden layer compensating for my mistake... -- Darkcanuck

All waves. I found that weighting firing waves higher decreases my general RM score, as well as my rumble score. -- Skilgannon

That makes perfect sense. My NN training scheme (which seems to be the key to getting good results) is based around learning firing waves, but a dash of all-wave learning helps a lot. I'll try to build a RM-tuned gun using all waves, we'll see how that works out. I tried adding some distance-since inputs to the NN but performance on this and the RM challenge didn't significantly change. -- Darkcanuck


Robo Home | TargetingChallenge2K7 | Changes | Preferences | AllPages
Edit text of this page | View other revisions
Last edited August 3, 2008 13:46 EST by wblv-ip-pcache-4-vif1.telkom-ipnet.co.za (diff)
Search: