Robo Home | TargetingChallenge | Changes | Preferences | AllPages

Difference (from prior major revision) (no other diffs)

Changed: 348c348,349
* It should. Maybe even against WaveSerpent. I'm not sure about any others ... --Simonton
* It should. Maybe even against WaveSerpent. I'm not sure about any others ... --Simonton
* I never focused on this version of DT, so it's tough to say. I don't think there's anything special about it - for the time, it was cream of the crop, of course, but comparing it to modern bots like Kev's Fortune series. Just get the Wave handling details as exact as possible (like surfing) and keep tuning your segments and weights. =) -- Voidious

/Constitution - /ReferenceBots - /HowTo - /FastLearning - /Results - /ResultsShort - /ResultsFastLearning - /ResultsChat - /Calculator
OMG!!! This is incredible!! ABC, I'd almost say that the rest of us need not bother trying to make our guns score more.... almost ;-) Congrats man! Just a question: I know you have a MultipleChoice PatternMatcher, but do you also do some segmenting? Distance to wall, Distance to enemy... that sort of stuff. btw, I added your average index in the Total column-- Vic

Not so fast - I'm working on my movement at the moment but I recon I still have a chance. Once I have the results I'll post them win or lose. -- Paul Evans

Thanks Vic, but first I need some more results to see how it compares. I did some tests with DT vs DT, it gets around the same results as Tron (very slightly better actually), but I don't know about the other types of movement... I like to call my targeting "pattern recognition", there is really no normal pattern matching involved. I'll make a detailed description soon, basically it looks for the 50 (ATM) most similar scans in the enemy's movement history and projects the following movements looking for patterns in the angles that would hit if he did exactly the same... I compare speed, distance travelled in the last 15/30/60/120 ticks, lateral/forward distance to wall, distance to me (1on1 only) and angle to me (1on1). One of the advantage is that I can "segment" on different variables on the fly, the data is allways present (30000 scans limit). -- ABC

To help see if I had ironed out the bugs in Tityus gun I ran it through the TargetingChallenge and it seems I have. Now what's holding this gun back is not bugs, but just plain limitations of the simple implementation of GuessFactorTargeting. The "real" Tityus compensates some for this, but I beleive those compensation tricks are against the rules of this challenge. Tityus obviously has a harder time hitting Cigaret than it has with DT. I beleive this has to do with Cigaret being more of a bullet dodger than DT and Tityus fires VirtualBullets every scan and not just along with real bullets. There's a TityusMega in the works that will hopefully deal with this better. -- PEZ

Jez@^&@ ! What brute force are you guys displaying here! Congrats iiley with your score over 90%. Simply awesome.. My ZoomTargeting level 1 gun (without MultipleChoice) cannot compete on this level. Ok, back to the TargetingLaboratory :-). btw, iiley, I've put your score on top (sorting by total index descending). -- Vic

Way to go iiley! Now, can't someone post results worse than mine? I don't dare test Marshmallow. =) -- PEZ

Just checking. iiley, those results were from a run with no previous stats, right? As I suggested somewhere in the constitution of this challenge I think we should have to indexes. One where previous stats are allowed and one where this is not allowed. -- PEZ

I remember i deleted the stats data before i run this league~;],and i will post Cigaret's result soon. -- iiley

Great results iiley, very close! ;). I think maybe we should post our bullet damage scores too, this index is maybe a bit to sensitive to the enemy surviving a lucky round or two... bullet damage is directly proportional to your hit rate, and is maxed at 100 points per round. I wont be able to test again till the end of the next week, I'll post the details then. -- ABC

I'll post the details too,but ...PEZ,can you give me a upload ID,sorry that i forgot the one you gived me before. -- iiley

Details posted. -- iiley

Hey, if 100 points per round is max bullet damage. Then why don't we calculate the index as:

100 * (damage_points / 50000)
That would remove all freak stuff like enemy getting ram damage points and whatever. -- PEZ

but,if enemy hit wall or hit you,then you can not get 100 points. -- iiley

Too bad. But I think we can live with it. -- PEZ

That would only mean that 100% would be unreachable and that's not a real problem I think. PEZ' suggestion would be much easier to calculate, especially since RoboLeague already averages your bullet damage. I quickly calculated that Cigaret would have 94.58, Tityus would have 87.83 and Mazer would have 87.13. I don't know why these scores are higher then the current indices. The drawback is that this new Index would not be compatible with the MovementChallenge Index. -- Vic

I, for one, could live with the non-compatibility too. Besides, we can always calculate the compatible index as well, for when comparisons are wanted. -- PEZ

I added PatternMatcherBot's results for reference purposes. Maybe someone can post results for Head On and Linear targeting. Maybe then it's a little more inviting for newbies to join the challenge. -- Vic

Yes,it very very easy to calculate,current method is so hard to calculate,i spend many minutes to calculate that~;].are we going to change to this method? -- iiley

ok with me -- Vic

I'd sayt it's settled then. I'll post my corrected indexes asap. About head-on and linear references. I have never coded a linear aim gun. But I can make a head-on gun quickly enough. I'll check back here in a few hours and id noone else have started with that reference then, I'll do it. -- PEZ

I'll make a linear ain gun(now),and then post its result(tomorrow).First i need to sleep now,zZZzzz~~ -- iiley

Sorry that i didnt upload the result of Linear gun this morning,I cannot be online in my room now as i had not pay for the network of this month.;[ -- iiley

I uploaded Gouldingis details. I don't have the time to calculat the indexes, but if someone else feels inclined, please do. I'll be away until tomoroow night. -- PEZ

Haha~~I am back being online now,and ... I uploaded the Linear's result,it calculated with the new method(100 * (damage_points / 50000)) -- iiley

Hey, PEZ, the link you put there is the same as the one for Tityus. -- Kawigi

Just added hacked FloodHT 0.8 for this and added him to the list, partially so I can gauge the dev version against it. Sadly, the dev version does a little worse against pretty much everyone, even the ones that it should theoretically do much better against (Cigaret and HTTC). With pre-saved data, FloodHT 0.8 does a bit better against some and a bit worse against some others (particularly the same ones the dev version should have done better against), probably due to the fact that it learns incorrectly against some opponents. -- Kawigi

Your index against DT is plain awesome. What index do you get against the new DT? I was in a real hurry when uploading the Gouldingi results last time. Not only did I provide the wrong link. I also miss-spelled (how's that spelt?) the result file name. I have corrected both now, but I was not aware of that Roboleague recreated the .robotcache when it starts up so Tron and DT actually fought back against the stationary Gouldingi. I'll fix the results soon enough and will come back with the indexes. However I wonder if this .robotcache issue is interesting in other aspects. What if ER, for instance, forgets the .robotcache from one run to another? Then all our efforts of saving data and keeping the files small and all would be very much in vain, wouldn't it? Expect for us "fishy" guys who prep our robot jars with saved data of course. Is this news only for me or is it just under certain circumstances this occurs? -- PEZ

This must not always happen, because it worked fine for me to modify the data directories and then start Roboleague. My guess is that it would do that if it had some good reason to update the robot database in general. Thanks about the DT index. It seems that I must be doing better than CigaretST against the new DT somehow, since Sedan does better against the new DT than BlestPain (in spite of Iiley's constantly improving movement). FloodMini firing waves only when he fires, does even slightly better than posted above against DT 1.91, and given plenty of saved data, gets over 80% against DT and better than 93% in general. I'm trying it with his normal gun (firing waves every scan) now, and I'll maybe post up some results later. -- Kawigi

I posted results from FloodMini under 4 condition - both with and without saved data. The ones with an asterisk, FloodMini is only firing waves when he fires a real bullet (with no asterisk means he's firing a wave every scan). One can see why his normal method is the one I used, and also why it's not completely the better method (as he learns falsely over time against several bots). On the Roboleague sheet, the first 10 rounds were with FloodMini only firing a wave when he fires a real bullet, starting with no data, and accumulating the data for the 10 seasons. Then I wiped the data again and made FloodMini constantly fire waves for seasons 11 through 20. The ones reported are seasons 1, 10, 11, and 20. In theory, FloodHT and FloodMini with no saved data (and no asterisk) should be only trivially different. I found one or two almost negligible differences in their segmentation algorithms that may be the reason for the 2% difference.

I also figured out Gouldingi's numbers. I noticed that almost all the scores posted have close to correct scores against most opponents, but unrealistically high scores posted against FunkyChicken. This is probably because they are scored using the old system and FunkyChicken hits the walls a lot, preventing people from getting as much damage on him themselves. I'm thinking these should be updated by their individual authors. I could do the updates myself given some time, except for Tron's, since the scores aren't posted. -- Kawigi

Thanks for the help with calculating Gouldingi's index. I'm positively surprised by those numbers. I'd say that AngularTargeting/Factored now officially is a reasonable targeting method. =) I recalculated Tityus' index to use the new formula. It raised the index against most bots, except Funky of course. I also separated the tables for results with saved data and those without. -- PEZ

BTW. Great score against TheArtOfWar Kawigi! -- PEZ

Thanks, lol. I was surprised at those results against Art, too. And separating the tables wasn't a bad idea, either. Maybe even seperating them by weight would be interesting ;-). Turns out FhqwhgadsMicro scores 89.something, as well, including 100% against Art. Unfortunately, it's going to take me all night to finish testing FunkyChicken... -- Kawigi

It would be interesting to see what the Gouldingi score would be if it used the segmentation of FloodMini. As it is now its gun is completely unsegmented. -- PEZ

I just added FhqwhgadsMicro and FunkyChicken to the index - competing for the title of "best nano index" and "best micro index". For both of them, I used season 2. Season 1 was when I forgot to make them fire only power 3. -- Kawigi

Updated Tron's results, this time with roboleague details. I'm very pleased to see that my targeting is on par with some of the best guessfactor statistical guns (where are DT's results?)! Got to start working on it again, I believe it still has some unexplored potential... -- ABC

You sure seem to be able to see something exploitable about Cigaret's movement that the other guns are missing! -- PEZ

Yes, I noticed that too, and also my low index against DT and Tron. It's strange because I almost never test against Cigaret (too slow). BulletPower/Survival? is definatly a major factor in a normal 1on1 fight. -- ABC

I don't suppose anyone would like to run a set of battles for DT for me please - just to save downloading all those bots and getting roboleague running again :) -- Paul Evans

I am actually very bored and feeling generous. EDIT: The repository is down, though, someone who already has the TC bots would have to run it. -- Kuuran

I posted DT's score. Poor Cigaret. -- Kuuran

Thanks Kuuran - nice to see DT does not depend only on it's movement. Don't suppose you want to run it again a couple of times with saved data - only if your still bored :) -- Paul Evans

Heh, I'll try to remember to run that overnight tonight, right now I'm working on refining Apoptygma's gun into a competitive micro. -- Kuuran

I knew you would win this one too Paul, congrats anyway :). -- ABC

I was not so confident - I was beginning to think it was only my movement that kept me on top - especially as the guns have not changed much this year. -- Paul Evans

Some strange observations: you seem to have trouble hiting Funkychiken, almost as much as Cigaret! Also, my movement doesn't look so bad now, Tron 1.5 is the 4th best at avoiding your guns. :) -- ABC

Actually, more than half of the current list has their index against FunkyChicken miscalculated :-p They should almost all be 8-10 points lower. The new way of calculating the index, though, (bullet damage/50000*100) gives slightly higher scores against most opponents and deflated scores against FunkyChicken, because it hits walls and takes its own energy from that. -- Kawigi

It's good this challange seems to at least somewhat be able to measure gun performance. Considering the relatively, against the movement challenge, small differences here I think this also serves to give a hint on the eternal question of MovementOrTageting?. All top-5 bots have really good movement and DT has clearly the best (especially since its guns identifies and exploits a weakness in the Cigaret movement). Movement is where the battle is decided as things stands now. Still, the best bot is best everywhere. I'm in awe that you can keep the #1 spot for month and month Paul. And keep improving DT.

Kawigi, if any of my bots have the Funky index wrong you would tell me I hope. -- PEZ

Mm... two editing conflicts on the same edit. Hooray. *glare*

I was going to point out that it was primarily matchers, not stat guns, which had the high index on FC, but if that's a result of the old formula then that explains it quite handily :P The hitrates on Aspid and Tron are somewhat surprising, though, I wonder if those movements taking enemy position into account less has something to do with it? -- Kuuran

CigaretST, Tron, Mazer and PatternMatcherBot are currently calculating their indexes incorrectly (on a quick look). FunkyChicken is convenient to have there for figuring this out, because the numbers will usually be +/- 1 or 2 from their current scores on every other opponent, but will be 8 or 10% too high against the Chicken if they use the old calculation system. -- Kawigi

I moved the erring bots into a table of their own. And whoever correct the indexes can move the bots back. -- PEZ

I didn't know we had decided for the bullet damage index... fixed. -- ABC

And this was obviously better for Tron than the other index. =) We should probably tidy up all TargetingChallenge pages so that things are a bit clearer than they are today. -- PEZ

Yes, much better :). I never really expected my strange gun idea to be so effective, I just tried my development version of Tron, with slightly improved movement, against iiley's bots, it wins against most of them, Cigaret 1.31 looses 40/60! I will release a 2.02 version this weekend just to celebrate that. :) -- ABC

"I was beginning to think it was only my movement that kept me on top - especially as the guns have not changed much this year." - IMO you have always been ahead of the times, both in movement and targeting. I still can't defeat SandboxLump (I'm finally getting closer though), the only pe.* bot I'm able to beat is SandboxMini... -- ABC

You forgot SandboxMicro? :) I started running DT with data tonight, and I'm heading to sleep shortly. I'll likely post results tomorrow. Total seasons will be 10 - I think DT will have learned his opponents given 4500 rounds plus the 500 round match that counts ;) -- Kuuran

Paul, I've run the league with data and uploaded the results to [http://www.geocities.com/kuuranca/DT_TC_Data.html]. The scores are pretty much identical with and without data (the one with data is actually a bit lower, but that's probably random chance). Looks like DT has finished learning quite early on in 500 rounds against these movements. I'll add the data to the table tomorrow morning if no one else has by then. -- Kuuran

I've just been reading elsewhere that RoboLeague clears data files, I don'y know if its true but it may explain the results. Thanks for the try, I owe you one. -- Paul Evans

But not on Kuuran's machine, so that shouldn't be the explanation here. DT is a fast learner and 500 rouns probably levels out any lower performance it might have the first 20 or so rounds. -- PEZ

DuelistMini's improved TC performance is due to better bullet power selection and better segmentation, including some segmentation designed to handle different bullet powers. --David Alves

Ah, I was thinking "this is too good to be true!".... In this challenge you must use only 3 power bullets, and fire them as soon as getGunHeat() == 0. -- ABC

Rats. Oh well, I'll change it after work today. :-p Thanks for letting me know. --David Alves

I added FhqwhgadsMini? (the dev kawigi.mini.Fhqwhgads) to the pile. -- Kawigi

Btw, I just remembered that DT in reference mode does not use saved data, that must be why Kuuran's tests came out the same regardless of the number of rounds... -- ABC

Oh, I think you're right. Thanks ABC, I was curious about that :) -- Kuuran

Good work with Musashi's PM gun! -- PEZ

Thnx, but not that good... I am trying something new (new for me, actually i think that ABC's Tron already do something like this) in v2.08, using three new parameters to rate the samples : distance from my bot, distance from the wall & distance to corner. I think that the bots probably move different based on how far they are from u,the wall and the corner. I don't know yet if it works to improve the choosen sample quality, have to test the older version (2.06) in TC and compare the results. The TC is a great idea to isolate and test guns. I even discovered a bug introduced in 2.08 by retrieving saved data between the matches - it scrambles the samples universe. -- Axe

Okay, I've updated the development version of Fractal and it gets just over 85% in the targeting challenge. I'm really impressed by this score. However, I'm looking to rebuild some of the smoothing I had from 0.32; this will take a little while though because I have to rewrite everything to support doubles instead of ints. I'm just worried that the data insertion may take too long. Hopefully this won't cause it to start skipping turns. After that I should see another small increase in its TargetingChallenge results, and then I think I'll start its movement. In the meantime, I think I'll run 0.32 through the TargetingChallenge for a comparison... -- Vuen

Phew; Fractal 0.32 got an index of only 80.8%. I was terrified that it would get higher. It's interesting then that 0.48 ranks so much lower than 0.32. I'm thinking now that it may have to do with the learning time. The whole concept behind 0.48's gun is that it doesn't learn slowly; if it does, the gun is useless. I'll run both bots through with a few seasons of only 35 rounds a piece rather than 500 to compare the results and see what's up; I may also repeat for 100 to test out learning time there. There's lots of potention in the gun framework to improve learning time, and so I'll use this kind of method to test improvements in the bot's learning time. -- Vuen

Strange; Fractal 0.32 got 78.7% over 35 rounds per pairing, while the development version got 82.1%. This means the development version not only has a better long-term gun, but also learns faster than 0.32. I can't understand why 0.48 is ranked so much lower; I didn't change much between 0.48 and the development version. In any case, because of this I'm packaging up the new version as 0.50 to see if it will improve its ranking. I really like this TargetingChallenge thing : ). The sad part is, it's 4:15 am, I have a 9:30 class, and I'm also working on an assignment... [edit] Oh yeah, the repository's down. *smacks forehead* Ah well, back to work... -- Vuen

Blessed are you (we!) who are sleepless, for yours is the eternal sleep. :-) -- Axe

Vuen and Wolfman seems to be quite busy... Things are going to be very hard when the repository come back, think that i am sleeping too much. Very nice job! -- Axe

I have a proposal: maybe would be interesting if we add a column with a short of description of the aiming system used by the bot in the TC results (maybe a table below, if the descriptions are not that short). For example: Musashi - Multiple choice PM. By making this it would be easier to avaliate the weakness and strongness of a particular aiming system against the reference bots, and perhaps give us tips of what is working and what is not. -- Axe

I call my gun AutomatedSegmentation; it's a form of segmented GuessFactorTargeting but with an arbitrary-dimensional framework and MultipleChoice segmentation depth reduction. It's supposed to allow it to learn faster than would an ordinary segmented GF gun. Maybe sometime I'll write up an explanation of it. : ) -- Vuen

Maybe I am just retarded, but I have as much grief with TargetingChallenge as with CurveFlatteningChallenge. If I adjust my gun to maximize TC score, I get over 90% in it, but 'practical scores' agains 6 top bots go down considerably. I blame this on limiting TC to power 3 bullets, which are rarely shot by my bot in real fight. Thing is, if your GF gun don't keep information on what direction enemy goes when being hit, you need to guess-adjust different power bullets offsets differently (power3 bullets have chance to hit enemy from its side in 2-3 timeticks, power 1 bullets can have this chance in 1-2 ticks). It can make a big difference when choosing right angle to hit. At least my 'real gun' gets only 88% or so in TC, but is in practice better then 'TC optimized' gun. -- Frakir

Interesting. I have experimented with bullet powers below 3 with VertiLeach and found that it lowers its performance on most lowranked enemies. I have attributed this to the fact that higher powered bullets stays at impact distance longer, which is what you are confirming here. The problem with allowing different pullet powers in the TC is that we couldn't figure out a way to calculate a comparable index. Solve this and I will be one of the first to vote for removing the BP3 restriction. And I know iiley would like this too since some of his bots have bullet power optimizers that are totally useless in the TC. -- PEZ

Little off topic, but this is how it started: I tried to write a function returning probability of bullet hit when shot by perfectly random gun. Simple: calculate how far can bot go one direction, how far in another, divide by twice the bots size adjusted for bullet angle... then real random gun gets some 1.1 times better hit rate then above 'theoretical'. This is due to bots being 2D and having some chance of moving into bullet form both sides (so bot effective hit size is larger then physical). I wonder if anyone wrote a gun properly adjusting for that.... -- Frakir

Quantums gun takes into account the size of the target, and there are a couple more things I can do to improve on it that I want to try out as well. --Wolfman

Hmph. Fractal's results in the 100 rounds section is only 86.88; I'm less than impressed with its current learning speed. It should be able to easily achieve a score of around 90. I'll hammer away at it tomorrow to see if I can get a better result. In the meantime, where can I get Tron 2.01? I'm currently testing against Tron 2.02, which isn't so bad, but the TCCalc doesn't calculate Tron's score for this :(. Oh and PEZ, don't fix this; it will force people to use the same versions of reference bots. Anyway, I was wondering if anyone wanted to participate in an even shorter league; something on the order of 35 rounds. You could make it make people run say 3 seasons of 35 rounds, no data saving, and then average the results across the 3 seasons to get a final result (thus reducing the 'chance' factor). This would be a measure of a bot's learning speed rather than long-term accuracy. Learning speed is what I'm going for with Fractal's gun, so this would be a better measure of its strength. I think it would be very interesting to compare the results between the 500 round and the 35 round challenge to see which bots learn quickly and which don't. -- Vuen

Cool. I can make a calculator for it. Let's make it 7 seasons. 245 rounds against each reference bot should iron out most of the randomness while not taking forever to run on my dated laptop. Download reference Tron from http:/robocode/uploads/pez/abc.Tron_2.01.jar -- PEZ

I modded TCCalc just slightly to make it useful for the /FastLearning challenge. -- PEZ

For reference I ran a challenger version of Griffon in the /FastLearning challenge. Not as good as Quantum but much better than GloomyDark. Which is a bit depressing since GD's gun is supposed to be evolved from Griffon's... -- PEZ

Thanks for the link to Tron 2.01 by the way :D. You may want to put a link directly on the TCCalc page so that people will know where to find it if they get a 0 score against Tron. -- Vuen

Results of Musashi v.2.08.8.TC (embryo of 2.09) added. Implemented a additional MultipleChoice inspired by ABC's TronsGun idea (get the n best samples (PM). Play the movie for all n in order to get the n points they predict. Calculate the n gun angles aiming the points, then take the one with the minor diff to the others). Amazing how it works well (not so surprising if we consider that its a ABC idea - Man, u r great!) -- Axe

BeerPoet results added. It uses HeadOnTargeting. -- PEZ

I've added Recrimpo 2.1's results, but i think they are unfair, because the result against Fhq.. was extremely low due to Fhq.. ramming me to death:

Rank Robot Score Survival Last surv. b. Bullet dmg. Bullet dmg. b. 2*Ram dmg. Ram dmg. b. #1st #2nd #3rd
1 Fhqwhgads 1.1TC 29171 23800 4760 0 0 546 65 477 23 0
2 Recrimpo 21502 1150 230 19666 456 0 0 24 476 0

Musashi's results posted 2003/11/18. Not so good, but not so bad. Is there a place to upload the details? -- Axe

I have just posted the results for a tweaked .dev version of Quantum. Im very pleased with the results!! I still have the memory problems, if I can sort em out perhaps I could improve the accuracy! --Wolfman

I also added the results for the development version of Fractal, and I am also very pleased. Hopefully the repository will be back up soon so we can compete again :) -- Vuen

I added the results for a bot-less gun i knocked up earlier. I'm hoping that by positing my feeble results i'll inspire myself to work a bit harder on it. Seems to get a bit of a pasting from funkychicken... --Brainfade

@Tango: F only rams you to death when you're already disabled as I understand it. This is only boosting his score, not lowering yours, and since only bullet damage portions of the score are looked at the index isn't affected. -- Kuuran

I added a still very raw stat gun based on code lifted from Apoptygma. The difference here being that the gun is leaving about half of a micro for movement and other things, and that this gun is meant to be competitive instead of a miniaturization showcase. GirlPower has tentatively named this robot "Muffin" o_0 -- Kuuran

When calculating the averall score do i have to use the correct single scores ( 78.52568 ) or rounded (78.53)??? However i added Idiot's score--deathcon

I may be interesting to add an additional column describing the type of gun (GF/PM/other) that is used. May be difficult to describe with a bot like SandboxDT though ;-) --Loki

Would it? It's GuessFactorTargeting if there ever was one. -- PEZ

Is that so? Didn't get that from the main TargetingChallenge page nor from these pages, but must admit i didn't read it all. Just thought that targetting is targetting, no matter what gun is used (posted results of Fenrir with PM-gun on /ResultsShort). It's about targetting here, not movement, so why not test different types of guns? --Loki

You can use whatever gun you fancy! =) I just meant that SandboxDT is the model of GuessFactorTargeting bots. Paul is the inventor of the that concept and I doubt he has even ever tried write a PM gun. -- PEZ

AFAIK DT has lots of GF guns, but nothing else (unless it's got head-on, linear and circular for good measure). -- Tango

I would tend to think that a GF gun can trivially become a head-on gun if it needs to be, and that he may well use guns that are guess factor guns based on linear and circular escape angles. -- Kawigi

ok, a lot of confusion (on my side), but what i originally sugested was to include a column in the table describing the type of gun a bot uses. That way the weaknesses and strength of a type of gun will become clear, not only which bot is the best at targetting. I have made a start. Can everyone check his bot and update when necessary? thanks. --Loki

For example, my conclusion (based on just a couple of bots ) is that PM-guns perform badly agains a bot like SandboxDT or Cigaret (resulting in a percentage smaller than roughly 65%), while doing reasonable well against other top bots like Tron. TAOW moves in a predictable way, making it an easy target for both PM and GF guns. --Loki

I think I finally got the bugs out of the gun. :-) -- Jamougha

Wow. *speechless* Now I guess all you need to do is get that AM right, and you'll be Good to Go! -- nano

It might be a bit over-segmented for the RoboRumble, but I guess I can fix that with some saved data in future... but generally, yes, that's the plan. ;-) -- Jamougha

Way to go! That DT score is sooooo cool. -- PEZ

Thanks :-) yes, this gun is now OK against the old DT movement. And from the fast learning results it looks like I haven't yet hit a segmentation limit, which is encouraging. -- Jamougha

I tweaked Aristocles gun some. Now it performs quite close to DTs score in the long run. The tweak doesn't affect its fast learning score though, so I might try different tweaks before release. -- PEZ

Not a bad result for a micro, regardless. :-) Keep tweaking, it would be very cool to have a micro gun beating DT! --Jamougha

I haven't tested, but chances are RaikoMicro does just that. You have better segmentation there. -- PEZ

Wow. Just... wow. Has anyone run a TC with the latest DT? it'd be awesome if you did better than it too ;-) --David Alves

I think any changes to DTs guns since 2.11 might not show in the TC. One change was about CribSheets$ the latest was anti movement adaption. It probably would show in the TargetingChallenge2K4. But feel free to test it. I think DT still has its behaviour properties for this purpose. If you test it, please update the above table. -- PEZ

He he. I guess a micro-gun can play too. =) -- PEZ

Wow, very cool! :-) And you are even beating my score against DT. -- Jamougha

=) I guess fixing bugs can pay! -- PEZ

Paul, I am running a full TC with DT 2.71 in roboleague right now. (Damn it's nice having a fast PC for a change :-) ) -- Jamougha

Jam, we can switch off if you like. I already did Tron and Funk, TAOW is in progress. Out of the remaining bots I'll do HTTC, Yngw, and Spar, you take the rest. --David Alves

OK, but when you both are finished with that you start your RR@H clients, right? =) I want help getting Aristocles 0.3.7 a rating quickly. -- PEZ

David (my mistake :-) ) Unfortunately I have done Funk, Aspid, Fhq, Yngwie, HTTC and Sparrow - btw how did Paul make a gun this slow? - but Roboleague is set up so I'll just let it run. That way I can upload the details sheet if someone has the webspace.

PEZ, I can't run RoboLeague from the university computer I suspect. I'll try it and see, though. :-) -- Jamougha

For comparison my full results gave

SandboxDT 2.71 Paul Evans 78.83 91.85 99.79 99.33 84.70 88.20 96.07 87.05 97.72 87.68 91.12

so it looks like 2.71 is worse at targeting than 2.11 :-\ -- Jamougha

Ahhh, I was worried first. But that could be withing the error marging I guess. Maybe it is as I said above, that there are no changes to the gun sinc 2.11 that would affect TC perormance. -- PEZ

Yes, the last thing we want is a bug in DT's gun. That would mean Paul could go and fix it... :-o -- Jamougha

I think the second score is close enough to 2.11 performance to show the gun has not deteriorated since 2.11 - It just goes to show what can be done with movement. Thanks for doing the test though. -- Paul

What can be done with movement and a strong gun that is. -- PEZ

does 80% against tron and 60 against dt2.41 mean a gf gun is working?--andrew

With no segmentation, that's about right. -- nano

I'm not sure. That would mean it scores on par with BeerPoet and Gouldingi. You'll have to run a real TC run (all 10 bots) to get a better picture. I would suspect a bug if I where you though. But run all bots and we should see. Unsegmented it should be somewhere around Nemo's score I guess, probably a bit better since Nemo's gun must sacrifice at least some accuracy to fit in a nano with movement. Segmented you should move closer to Racrimpo? I guess. WIth any fancy segmentation you should be able to beat the old Tityus score there. Because that gun was buggy (I'm still thankful to Jamougha for finding the bug!). -- PEZ

What sort of gun is it that stores patterns for each segment and then matches those patterns to the latest movement to determine what guessfactor to use? -- Alcatraz

I'd say it's a PatternMatchingGuessFactorTargeting? gun. =) Have you succeeded in creating one? If so, I envy you big!-- PEZ

i guess it's not working because the declaration for the stats array looks like this: static int[][][][][] aimFactors = new int[2][3][3][5][25]; maybe my segments aren't working somehow...--andrew

i get 84% against sparrow also....damn this--andrew

I think that Paul's hiseg scores basically show that he hasn't figured out the optimal way to switch between VG's, either. -- Kawigi

How do they show that, kawigi? --David Alves

If the hiseg results are better than the normal results, even over long battles, it means he isn't dynamically switching to hiseg with VGs. -- Tango

hiseg mode is disabled by default because it makes DT run really slowly and consume tons of memory. In other words, Paul turned it off because he's having pity on our computers, not because he's unaware that hiseg mode is more effective... --David Alves

Quite. I was just trying to interpret Kawigi. -- Tango

Jamougha and / or PEZ: Any tips on how to raise my TC score vs. DT 1.91? I seem to be able to hit every one else pretty well but I can't figure out how you guys are getting 86% vs. DT, I can't do better than 82%. :-p --David Alves

Are you using a time since last stationary segment? -- Kuuran

I think it might be worth trying a bit higher guess factor granularity for your waves. DT stays far away if it can. -- PEZ

@Kuuran I'm using every segment known to man, and then some. :-) --David Alves

David, I've found time since last stationary is less effective against DT in general than time since last acceleration. The reverse is true for some bots, though, and there are other time-since segments which are more effective against yet others. Try swapping and see what it does to your TC scores. -- Jamougha

Impressive results for Brachistochrone... What sort of bot is this? --Alcatraz

I remember reading somewhere that he was making a GF gun. Unless he has come up with something new and exceptional my guess would be that he is firing with other than power 3 bullets, because in my cynical little mind i doubt that a GF gun can be made that beats Raiko, DT et al by such a large margin... --Brainfade

What did u feed with Pugilist's gun, PEZ? Quite impressive that results! -- Axe

I'm getting this message in the battles against DT 1.91 in refrence mode:

GE> pe.SandboxDT 1.91 still has not started after 7200 ms... giving up.

Is there any workaround? I don't remember it happening before... And it makes me lose around 20 rounds worth of bullet damage.

-- ABC

Strange. I seem to recognize that problem from before... That was a later DT though. I think it was slow at reading it's data and thus slow at starting. Are you sure you have DT in reference mode there? -- PEZ

I'm sure. It only happens in RoboLeague. I seem to have found a solution, I included the "-Xmx512M" flag in the "user.cmdline" line of RoboLeague's "launcher.properties" file. It's a bit strange that it's DT that gets disabled when I use a lot of memory, got to optimise my gun's memory usage... -- ABC

Yes, of course you're sure. I realized the question was stupid when I had posted it. =) But DT is a bit unstable really. My RR@Home client only crashes about once a week (it's running 24/7) and none of these crashes is without DT being one of the opponents. But please don't hesitate to do something about your bots resource hunger. =) Always when my interactive test runs goes to a halt I check what battles the rr@h client is running and it's Cigaret or one of your bots then. The real killer is when it's Cigaret vs one of your bots. =) -- PEZ

Strange, all my RR@Home client crashes happen when SilverSurfer is fighting, I've never seen it crash when my bots are fighting... :) -- ABC

But maybe that's because your client get the shakes when SilverSurfing runs. =) -- PEZ


This may be a bit off topic, but I'm wondering if any of the higher-ups on that list keep a record of their hits/total_shots ratio. I'm just curious about what type of percentages a good gun produces. (I think that this percentage is only partially related to average bullet damage.) --wcsv

Something I've been wondering about Fhqwhagads for a long time. Basically any decent GF gun can get 95 or 96% on this bot, but the top 3 have a 2-3% separation. What is this due too? -- Alcatraz

Frankly, I have no idea, as I don't usually run the original TC as a benchmark. Dookious and CC are OpenSource, of course, if you want to take a look yourself. =) I'd have to take a closer look at Fhqwhagads' movement to have a better idea, but the first place I'd look would be wall-distance and reverse wall-distance segmentation. -- Voidious

Ouch, those are impressive scores for a mini. Komarious really sets a new standard. -- GrubbmGait

I'll be posting some Ugluk scores later today, I think. That slapping sound you've been hearing is Ugluk spanking GrubbmGrb and GrypRepetyf. Hard. -- Martin

So Voidious, PEX, Mue, and maybe even Alcatraz. Whats the secret to getting a 90% vs DT? --Chase-san

Robo Home | TargetingChallenge | Changes | Preferences | AllPages
Edit text of this page | View other revisions
Last edited September 17, 2007 17:08 EST by Voidious (diff)