Archived talk:Ascendant 200502
|
This looks to be a strong bot! It beat RaikoMX and Silversurfer and has a rating of 1945 after 171 matches. That's a >100 point jump from Hyperion. Did you only replace its movement with WaveSurfing or did you change more? Well done! --Vic
Thanks :-). I'm indeed quite pleased with the performance (although there could still happen alot until the 500 battles are reached). And yes, i only changed the movement to see it's influence on the results (i already suspected that Hyperion's movement prevented it from rising further - now i'm sure of that). --Mue
Hell of a strong bot! Congrats, man! If that´s your prime version, things r surelly going to get hot! -- Axe
Congrats from my side too. But i am not going to loose my place as best german robocoder without a fight ;D --deathcon
Hehe thanks :-). Its just my prime version of WaveSurfing, everything else is taken from Hyperion. And yeah, i noticed that Shiva just jumped in front of Ascendant. That little bugfix you mentioned on its page must have helped a lot. Lets see if i find a way to get rid of some of the many problem bots... --Mue
Yeah it was that bug i hoped to find. You know that: you just finished a bot and it performs bad and there is that hope that it is a huge bug that makes your bot perform so bad. Because a bug is easier to fix than a bug to improve ... ! good fight ... --deathcon
Sure i know that hope :-) This time though it looks like i'll have to do some thinking on how to improve (no obvious bug found yet). --Mue
Good luck. The run for 2k begins :D --deathcon
Aim higher dudes! Go for the crown! -- PEZ
Looks to me like you have already started running without me :-). Ah well, i'll probably need more time anyway, since i plan to stick to movement changes until the next release. Sadly i'm having some difficulties to find a change that makes Ascendant do better vs the problem bots i'm testing with. To make things worse i get lots of ideas on how to improve the guns while experimenting with movement... --Mue
Take notes of these ideas. One day u will need them badly. --deathcon
Sure i did and i'm really looking forward to experiment with the guns. After a week of movement tuning i get the feeling that its way more fun to blast the problem bots to pieces than to trying to evade their bullets ;-) - at least when all my movement changes show if any only minor effects in the tests. Now i let the rumble figure out whether the changes where any good at all. --Mue
Ascendant 0.5 likes a close fight. It's the only top bot that likes that I think. Excellent to add to your test bed for tuning your close combat movement. I corrected a huge problem CassisClay had there with the help of Ascendant.
Mue, I haven't tested 0.5.1 yet. But if you haven't experimented with it much, try evading more from your opponent. It might boost your performance more than you think.
-- PEZ
Hmmm, it doesn't seem like you need to improve your movement too much actually. Ascendant is a slayer of strong bots. Could be your guns too. Maybe VirtualGuns isn't such a bad idea against surfers? But I think maybe your surfing style is extra vulnerable to PM guns? I wouldn't have guessed from watching it fight, since all those stops and go's should be tricky for a PM gun to figure I thought. But why should Robocode all of a sudden be figurable? =) Great bot you have here anyway. It is definately built from material giving it #1 potential. Try plug the .p gun in and see if you might need to think differently with those guns of yours. -- PEZ
Thanks for feedback :-). Actually i already tested with keeping farer away from the opponent, but this performs worse vs almost all bots of my testbed. That could very well indicate that my guns are simply not very good at long distances. So i will test that again after improving the guns. And you are definitely right that 0.5 had a problem with pattern matchers, especially when they use absolute bearings or guess factors for matching. Thats why i introduced another style of surfing (similar to TrueSurf) and switch between the styles when the opponents hitratio becomes scary. But well, it does only work vs some of the problem bots and even then not that good.
And for testing i already took the CassiusClay movement to see how it deals with my testbed. It was considerably better vs the problem bots, but lost some points to others making it overall about as good as my current movement (for the testbed of course, could be different in the rumble). I'll probably run a similar test with it's gun just to see, what real top bots can achieve with their guns. I really appreciate having such a plugable movement/gun of a top bot at hand, thank you PEZ :-)
Well, obviously the movement effort was not wasted like i feared since it got me 20 points, which is definitely more than i expected. --Mue
The movement is awesome I think. Can you give some more details on which bots our two movements differs in performance in your test bed? That might give me some clues on what to try next. CC's raise to royalty doesn't exactly go on with the speed I wanted it to. =) -- PEZ
Of course i can :-). Though i have to admit that i didn't note the version of CassiusClay i took the movement from. It was certainly not the most recent one (which changes twice a day :-) but probably 1.0. |opponent | A's movement | CC's movement | |syl.Centipede | 80% | 70% | |nic.Nicator | 89% | 81% | |dans.Cinnamon | 85% | 93% |
The numbers are the averages of the score-percentages Ascendant got in 15 battles with 35 rounds each (thus a higher number means better :-). All other differences were within the margin of error, since results may very well vary by about one or two percents. I just noticed that my last movement change almost closed the gap to CassiusClays movement concerning the real problem bots i tested with (which is why they dont appear in the table above), i'm quite happy about that :-)
And btw, i might have noticed a small error in CC's movement. When reacting to BulletHitBullet-events, it registers a hit using event.getBullet(), but i'm rather certain that this will return your own bullet while getHitBullet() is the one you want. Nothing big but... --Mue
Thanks! I'll have to figure about what the table above means now. =) I'll correct that bug. Though I only use the coordinates of the bullet, which I guess makes it less important which one I use. But it does add confusion that I don't need. =) -- PEZ
OK, i didn't look that deep into the code and thought you were using bullet.getHeading() to get the guess factor the bullet was fired at.
Hm well, i tried to explain what the numbers mean, but my english might have confused the matter :-). The first row: I ran 15 battles (with 35 rounds each) Ascendant (with it's own movement) vs Centipede and another 15 battles Ascendant (with CC's movement) vs Centipede. In the first setup Ascendant gets 80% of the score (averaged over the 15 battles), the second setup results in 70% for Ascendant. I hope this clears things a bit. --Mue
Well, your english is perfectly fine. I understood the figures as such. =) -- PEZ
Really impressive your new version. And when i looked at it detailed ranking table i wondered: it shares exactly the same problem bots as Shiva : Fenrir, Booring,tm.Yuugao, tobe.Fusion, the tron family, GlowBlow, pi.Dark, mld.Moebius, Smoke, Aspid and so on ... . I am not sure how this comes, but it might come from our similar way of surfing waves ( much like TrueSurfing ). --deathcon
It's the same problem bots as all wave surfers have I think. -- PEZ
Wow! Ascendant looks really good now. I think that loss against Cigaret is partly due to bad luck and partly due to your gun not being strong enough. You probably have some 30-40 points to collect by fixing your canon. -- PEZ
Well, vs the cx-bots i always need some luck, i still have to figure out how to beat them consistently. The problem with the guns is that i already tried to fix them and i dont see me gaining 30 points by just tweaking them further. Perhaps i'll release a version without the virtual guns stuff to see whether there are some points hidden... --Mue
All of those bots are pattern matchers. Most are slightly atypical ones at that, Tron uses it's own famous matcher, Aspid and Moebius use LateralVelocity, GlowBlow uses a very well designed heading/velocity, I believe the same is true for Smoke. It makes sense that plugging a stat gun into your movement would benefit you in all areas, but it also makes sense that pattern matchers would find ways to hit you that it doesn't compensate for as effectively. Same goes for adapting to the last hit at guess factor, since the matcher can track several movement patterns and switch immediately rather than wait for a new guess factor to fill up. Has anyone ever tried putting a fast pattern matcher on their adaptive bot, converting it's output points to their appropriate guess factors, and inserting those points along with wave data into the self stat array? -- Kuuran
Hail! 2K!!! This is a historical moment. Congrats! -- Axe
Way cool. I have seen this coming and now you did it, congrats from me. --deathcon
Thank you, it certainly took me long enough to reach it :-). Well, finally a member of The2000Club (thanks for adding me deathcon), i'm quite happy now :-)
@ Kuuran: Sounds like an idea worth trying. I'll probably test a bit around with it as soon as i'm done with gun tweaking (which might be very soon). --Mue
Congrats!! I was hoping it would be next but it is nice to see someone else do it. That means it is still possible :) -- jim
It's a bit belated, but... Welcome to the void land between 2020 and 2050 points! Good to clean out some old assumptions from your gun, huh? =) -- PEZ
Hehe thanks :-). I just had to try some things in the rumble, since my tests lack the precision necessary to assess the value of some tweaks. The good thing is that with the release today, i've finally worked off my todo-list concerning the gun. Thats maybe a good time to turn to the movement, which has been unchanged since 0.6.1. Anyway, i'm very impressed that some bots score still 40 points better than Ascendant now, when i even consider gaining 10 points very hard ... --Mue
Please define "a more precise guess factor calculation". Thanks! -- jim
In the guess factors calculation i used before, i assumed that the target can instantly go forward/backward at velocity 8/-8. This is an approximation, since the target cannot do that (especially going backward with velocity -8). It may even be that none of the negative guess factors are reachable at all, if the target is very near. So i tried to determine the angles corresponging to the guess factors 1 and -1 by iterating the number of ticks until the bullet hits while applying the acceleration/deceleration rules. I still assumed that that target can instantly go perpendicular to you, which is not that big an approximation.
Anyway, since the change did not show any effect in my tests i just released it, to see how it does in the rumble. As you can see, it did not really show there either. I even lost some points, but i consider that to be just a fluctuation. Its probably because i already segmented on velocity and distance, so that the actual bins already reflected the unreachability (does that word exist?) of certain guess factors. So all in all, not worth the effort... --Mue
Why would accounting for bot width in the -1 and 1 GFs be important? I have reasoned that it sorts itself out since you use the same calculation when collecting the data as you do when you target with it. One more question. Is that -1 GF the unreachable one or have you also accounted for reachability first? -- PEZ
I reverted to not account for reachability (i tend to remove additional computations that do not pay), so the GFs 1 and -1 are just the normal ones. So why accounting for botwidth in GF's 1 and -1?:
I noticed that my guess factor calculation when collecting data sometimes produced GFs slightly above 1. I reasoned that this happens because i actually determine a guess factor range that would have hit (using the bot width). So if the target went to guess factor x, i updated all bins corresponding to guess factors between x - margin and x + margin (margin being determined from bot-width and distance). And if x is close to 1, x + margin may well be beyond that. Thats why GFs close to 1 resulted in fewer bin-updates than lower GFs. I thought that this asymmetry might somehow influence the performance, so i just expanded the GFs 1 and -1, making sure that all possible guess factor ranges are subsets of the interval [-1,1]. I was quite surprised that this lowered the performance of the bot, but well, its not the first time the rumble does not behave like i expect it to :-) --Mue
For CC updating anything else than only the single GF bin converted from angleOffset results in a performance loss. Have you tried not bothering about bot width at all? -- PEZ
No, not yet. But i'll try that now :-) --Mue
FYI. CC suffers a perfomance loss if it considers bot width in the surfing too. Though not quite as much. -- PEZ
Wow, it's only 434 battles yet, but it's not going to recover 25 points. This is interesting! -- PEZ
Well, i would have had difficulties explaining why the bot performs better when not accounting for bot width. Its probably this way because Ascendant has a lot of guess factor bins and usually fights at medium distance (around 400 pixels) (yeah, first thing i do when turning to movement changes is trying to increase this distance). --Mue
And if you try using 27 guess factors in your gun and otherwise keep the 0.7.6 config? -- PEZ
Well, it doesnt hurt to try i guess. Released 0.7.6b that uses only 27 guess factor bins. --Mue
No it definately doesn't hurt trying! I wonder if I have lots of points to gain by using many more factors and bother with bot width... How many factors do you use? Do you update all covered bins equally much or do you weigh for distance to the middle bin? -- PEZ
I use 47 bins and yes i update all covered bins equally. Though i never thought about taking the distance to the middle bin into account => i'll experiment with that a bit :-) --Mue
That's what I tried when I dealt with bot width. Never tried updating it flatly though => guess what. =) -- PEZ
"rolling movement stats slightly slower when hit and faster when not hit"... What does that mean? Do you use the same buffer for flattening and hit avoidance? -- PEZ
Yeah, i use the same buffer. When hit the stats are rolled very fast (slightly slower is still fast) to make it almost impossible to hit twice in a row using the same guess factor. And if an opponent uses an advanced gun and a bullet misses, stats are also rolled (of cause slower than when hit) and a 'minor hit' is applied to the GF Ascendant was at. So if the opponent seldom hits this results in some kind of flattening i guess. --Mue
Do you protect the flattening from happening if the enemy virtually never hits? Like with head-on targeters. You probably never should allow flattening against those. This could be why you don't shut them out like many other surfers do. CassiusClay checks to see if it is being hit often enough before it switches on the flattening. I think this gains me quite a few points. -- PEZ
Actually i only start flattening when the opponent has already hit Ascendant several times at different guess factors. Do you have some specific bots in mind, that do better than they should? I mean, i know of cause that there are many bots CassiusClay does better against, but i wasn't aware that there are bots with really simple targeting methods among these. --Mue
I think there are plenty. If you have a tabbed web browser. Try opening the LRP graph of two bots in different tabs. Then flipping back and forth between the two graphs gives you a pretty good picture. Of course 35 rounds give room for a lot of randomness, but the overall picture is usefull I think. If you want to pep yourself to great deeds do this with Shadow and/or RMX against your bot... =) -- PEZ
Ok i see what you mean. Maybe i should make sure that my detection of when to start flattening works as desired. --Mue
Good idea. From my tests EnemyWave's are only so exact. I register hits on quite a few guess factors even from a head-on targeter. It sounds unecessary too. If you are hit more than, say, once every round you are dealing with a gun which doesn't shoot the same guess factor every time. At least if you measure hits above a certain distance. -- PEZ
There must be a horrible bug in 0.8.2. It looses to some bots with only some 5% of the score! -- PEZ
Thanks for watching. It definitly looks like a bug, so i recalled that version. Too bad that i'm unable to reproduce these results. Could you please check whether there is a Ascendant.log file in Ascendants data directory on your client? Only if it was running this night of course. Otherwise i'm out of ideas... --Mue
No data directory created there. I'll keep an eye on it. -- PEZ
Ok, didnt really expect one. Thanks anyway. --Mue
Ascendants entry in the RR@H table is wrong ( it links to the old 0.8.1 version). I might be the only one who noticed, because i dont have this versions because i was off some time, and so havent already downloaded. Could u upload the 0.8.1 version to repository once more or are u going to release 0.8.3 ;) ??? --deathcon
Ah, it never occured to me that i should upload the old version again when switching back in the rumble. Next time i'll keep that in mind. Anyway, i just released a new version which solves the issue (no 0.8.3 though since i still want to know how 0.8.2 performs). --Mue
Strong bot that 0.8.4 one. I see you still have some room for improvement when it comes to perfecting your results against simple targeters. 2035 points and that improvement opportunity. Promising! -- PEZ
Thanks. I'll probably look next at some of these simple targeters, maybe i can figure out what's going wrong there and finally claim #4 in the rumble :-) --Mue
I think performance against the really simple targeters are just about ironing out the last flaws in the surfing. I've found small and big bugs in CC lately. And made the calculations more and more precise and with steady steps I've closed the gap on how Shadow and RMX handles those opponents. Not that it has payed all that much in rating points. I'm not even sure I've gained any points at all actually... But at least I haven't lost any points and the more exact and solid infrastructure is easier and more intuitive to build on. Run 1000 rounds against Barracuda and check the results. I propose a WaveSurfing/BarracudaChallenge for this. Should help many of us get our surfing act together I think. -- PEZ
Ok, i'll run some tests vs Barracuda and enter the score there when i'm home again. I already found and fixed some issues with bullet detection by testing with a pure head on targeter. --Mue
Wow. 0.9 is going good. Circling the current enemy origin you say... I'll try that too. -- PEZ
- And i thought everybody is doing that already... --Mue
Wow again. 2063 with 600+ battles. Huge congrats dude! -- PEZ
Thanks, i'm impressed myself :-). My tests showed that it does better vs some simple targeters, but i did not expect it to improve that much. But well, finally Ascendant lives up to its name... --Mue
And what changes did you make besides the choice of orbit center? -- PEZ
Wow! Great difference! Amazing version! -- Axe
Thanks again, i'm already looking forward to upcoming releases, when every minor change i do will drop the rating :-)
@PEZ: Actually Ascendant uses two different surfing styles and switches between them when the opponents hit ratio increases beyond a threshold. I changed both styles to orbit the opponents position, but only one of them did profit from this (the one being similar to TrueSurf). Since this style did now considerably better vs some simple targeters i made it the primary surfing style (before it was only secondary, so that it would have never come to use vs simple targeters, since these dont hit hard enough to force a switch). So the improvement could come from orbiting the opponents position, or just from surfing style number two being better than the other one in dealing with simple guns (but most probably from a combination of both). --Mue
OK. Thanks! I get better results from orbiting the wave center as long as the enemy isn't close. Since you tried two things in the same release, maybe you should try this new surfstyle preference combined with orbiting wave centers? That way you (and I hehe) will now better what pays or not. -- PEZ
Woohoo! What a great result! On par with RaikoMX! Congratulations, Mue! --Vic
Ok, just made a release that orbits the bullets origin and is apart from that identical to 0.9. The next step will be to consider distance when deciding what point to orbit. --Mue
Obviously blindly using the wave origin isn't as good as orbiting current enemy location (for Ascendent at least). I'm following this experiment with great interest! -- PEZ
Amazing how Ascendant ranks 10 points lower with that change, are you sure you don't use the "wrong" orbit origin in the future position prediction? -- ABC
Yes i'm fairly certain. Maybe its not that important which point to orbit if the bot uses clever/sophisticated distance control, which is definitely not the case for Ascendant. --Mue
The enemy position is not predictable like the waves origins (they are stationary).Tbh, I thought you would gain points by changing to wave orbiting. Maybe the constant decision changing is good for Ascendant because it adds a random factor? -- ABC
ABC, are you saying you are not taking a surfing descision every tick? On the subject of orbit origin I have thought about it now and then. Intuitively orbiting the current enemy location should be better. Maybe orbiting the wave origin maximizes your escape envelope, but from a surfing point of view that's not too important. And surfing the wave origin at close ranges could get you into trouble I think. -- PEZ
I take a descision every tick, but my future tranjectory (assuming constant velocity) stays the same until a new enemy bullet is fired. If you orbit the enemy's position your projected tranjectory changes every tick, unless the enemy is stationary. Shadow's first wavesurfing versions also orbited the enemy position, I'll do some tests... -- ABC
Well, thats the reason i originally decided to orbit the bullet origin. Perhaps i should rerelease 0.9 just to check whether its rating was just a fluke. I still dont know how precise the ranking is. On RoboRumble/BattlingSystem PEZ mentioned that the rating of different releases of the same bot (same code) may differ by up to 10 points. --Mue
I think those situations are rare. Lately I haven't noticed differences that large. +/- 5 points I think is reasonable to assume. This is based on the differently versioned, but identical DarkHallow's that has been released lateley. It wasn't fluke Mue. You have a mighty strong bot. Make sure to install the strongest version before tomorrow night when the rankings are archived. Then after 00:00 CET you can start experimenting again. =) -- PEZ
No doubt, Ascendant is mighty strong. I made some tests with Shadow orbiting the enemy position, it performs slightly worse, just like I expected. -- ABC
And do you orbit the wave center even if you are close to the enemy? -- PEZ
Yep. Except when there are no bullets aproaching. -- ABC
Cool. That was exactly what Ascendent 0.9b tried i think. Which means this is yet another thing where it depends on your particular surfing style / implementation what's best. Everybody take note. Question every practice. Even those used by the very top bots. It might not apply to your bot. -- PEZ
Wow Mue. Now I really must find some time to work with my bot. 2067 points, that's awesome. My congratulations! -- PEZ
U r really a good bot name chooser! Congrats! Probably i should rename SS to Descent.:) -- Axe
Yeah, thank you :-). Just 4 points below CassiusClay; the crown has come in reach. Though i have to admit, that i have no idea right now where to get these points from ... next on my todo-list is some refactoring. So you see: no need to panic PEZ ;-). And yeah, i like the name too :-) --Mue
Well, I have no clear clue at what to work with either. It would be intersting to know if our bots are on par on both movement and gunning. If not then obviously we should be working on whatever of those two are the weaker part. What do you think? Been boosting your gun in secrecy? Or is it the same as any version of your gun where you tried using mine? If you ever did that. I can't recall... -- PEZ
Well, i never released a version that used the Bee gun. But i consider the RRGunChallenge to be a very reliable way of testing gun performance. Ascendant uses basicly the same gun as AscendantRRGC in that challenge (maybe with some minor changes i dont remember rigth now), so it seems that your gun does a little better (judging by BeeRRGC). I think that the ranking difference between CC and Ascendant could very well come solely from this difference in gun performance. --Mue
Which would help you with your todo-list. But not me. =) -- PEZ
Indeed :-). Redoing the gun is on my todo-list as well now. --Mue
Only 1 point behind now! -- PEZ
I guess Ascendant + BeeWT would be the absolute BOMB if such a combination would be built :-) --Vic
Yeah. It would be very interesting to try! -- PEZ
Well, i believe that the difference in the rankings of CC and Ascendant comes from targeting (which implies that the movements are about equally efficient). So i dont see any reason why Ascendant + BeeWT should do better than SilverFistWT or CC + BeeWT (the latter referring to the CC-version that got 2070+ of course). --Mue
Subsequentionaly congratulations mue for the big jump Ascendant did. Hope to see you up the soon :D
IMHO SilverSurfers movement + Shadows gun could be killer.
--deathcon
Hehe thanks :-). And according to the RRGunChallenge there is no big difference between the performances of Shadows gun and of Bee. Furthermore the talk above was about WikiTargeting, which does only work for guess factor targeting (well, maybe for some other forms of statistical targeting too) and is thus not compatible with Shadows gun. --Mue
I don't know ... I have a feeling you have surpassed SS and CC with Ascendant movement-wise. Of course it would also be interesting to see CC+BeeWT after PEZ's recent updates. --Vic
Well, maybe i'll make a release that uses the Bee gun (but without WT, since i'm still not sure whether i like this kind of massive pre-saved data). Thats probably the easiest way to compare the movements of CC and A. --Mue
Did my CC test with WT reach 2070+? I don't recall actually. In any case, the BeeWT gun is by far stronger than both Bee itself and Shadow's gun. Only against the RR@H population of course, but still. -- PEZ
Mue. I just noticed Ascendant always fires power 2.0. Have you considered changing this to 1.9 instead? It can make a big difference is my experience. -- PEZ
Yep, that happens to be what i'm just experimenting with :-). I'm not yet sure though whether using 1.9 for all distances works best for me (the tests tell that its definately not bad but i still want to test some alternatives). --Mue
How do you test a think like that? You need to battle the majority of the RR@H population to get a valid answer I think. -- PEZ
BTW. The CC test bot with BeeWT gun reached 2068 rating points. -- PEZ
Well i simply run lots of battles vs my testbed and assume that better results in these battles translate into more rating points. These predictions are not perfect of course, but i get a pretty good idea of whether a specific change is worth trying (by building a release for the rumble) or not. --Mue
So you decided on testing 1.9 instead of 2.0? Now I think my crown is in danger... I think lots of points are lost in that tiny space between 1.9 and 2.0. This will be very interesting to follow! -- PEZ
Yeah, though my test did not really show an improvement, so i'd rather be surprised if it does so much better now. --Mue
You must have a pretty amzing test bed to catch improvements of a few rating points with it. But I'm hoping the change will not make a difference for ya! =) -- PEZ
- Well, its not that good. My predictions usually involve a lot of guess work, and are thus sometimes proven wrong by the rumble :-) --Mue
2070 with 641 battles! I've reinserted my champ version (CassiusClay 1.9.9.09) in order to settle the crown issue once your new rating is stable. -- PEZ
Looks to me like a new King ascended to the throne, congratulations Mue! -- ABC
Yeah! Funny it should stabalize as I wrote that. Congrats Mue! -- PEZ
Thank you :-). But i have to admit that i'm a bit confused now. I seem to remember a version of CC that scored 2070+ points too, and according to CCs History that was version 1.9.9.05. So why didnt you reinsert that version? Well, i shouldnt complain i guess, its definitely nice to be the king even if its just for a short time :-) --Mue
CC version 1.9.9.09 is about the same as .05. Just a saner wave handling in the gun in .09. And .09 has also been seen with 2070 points in the table. Both these bots have got a lot of battles and many of them focused against a few enemies so their ratings are probably a bit scewed. In any case. No one of them can compete with Ascendant at the moment. The crown is yours man! -- PEZ
So, Long live Mue! Congrats man! -- Axe
I just tested Ascendant 0.9.2, it kills Shadow by a good margin, great work. It's nice to have a King that beats me for a change, welcome to my priority testbed. ;) -- ABC
Thank you, i'm honored... and i figure that this wont last (Ascendant beating Shadow i mean :-). For your information: Shadow has been long time member of my testbed and still beats Ascendant in short battles (35 rounds) by taking about 52% of the score on average (to be accurate: Shadow 3.31 does, not sure about the newer versions). --Mue
Congratulations Mue! Nice position. But now that Ascendant has reached the top it may need to be renamed! --Loki
Nah! It can still ascend in the rating space. -- PEZ
Just for everyones info. I have tried many, many bullet power schemes and not found any that works better than firing near 1.9. Lowering bullet power further for high distances might not hurt. But it sure doesn't help either. You're just trading bullet damage points for higher accuracy / survival anyway. Nota bene, I haven't experimented much with battle fields other than the 800x600 one. You might not visit higher distances enough for this kind of scheme to pay off. -- PEZ
Well, when a change does not hurt point-wise and increases survival i tend to keep it. But you are probably right that right now these higher distances are not visited very often. I think i'll play around with different distances now, maybe there are some points hidden too. --Mue
But I'm not sure you really get higher survival through the added complexity. That can probably be tested against a good TestBed... What's your current distancing strategy btw? -- PEZ
Well the new bullet power scheme did better vs my testbed, thats why i released it. Doing better vs my testbed usually means being stronger vs strong bots. Right now Ascendant is undefeated (yeah i know that this has probably more to do with luck than with the change). That is the main reason i did not switch back to bullet power 1.9 immediately :-)
Current distance control is rather simple: if closer than 300, increase distance; if farer away than 400 close in. I'm currently trying to increase the average fighting distance. --Mue
You have to have extremely accurate surfing to fight at distances around 350! Btw. I don't know if CC is part of your testbed, butI just tried 500 rounds with A 0.9.2 and CC 1.9.9.33 and CC was utterly crushed! I couldn't bring myself to check how the score was shared, but it must have been 65%+ for Ascendant. -- PEZ
Well, CC is currently not part of my testbed. But even if it were i wouldnt have noticed that since i usually run lots of short battles and calculate the average score share (and i'm sure the result would be much more even then). But nice to know that A holds its own in longer battles too :-) --Mue
Not using my RoboLeague/ScoreAverageAddOn? -- PEZ
Mue, congrats! #1 in RR and the PL. No lost battles! Truly amazing. --Vic
Thanks :-)
@PEZ: No, i dont even use RoboLeague. I wrote a small class to run the battles and calculate all interesting stuff when i didnt knew about RoboLeague. --Mue
Cool. Something you can share? I'd like a small class that I can hack on to automate stuff the way I want them. -- PEZ
Sure, why not. I have to admit though that its actually three files you need. I uploaded them to
http://informatik.hu-berlin.de/~ueckerdt/runbattles.zip
The important class is develop.test.RunBattles. Challengers are the bots you want to test, and opponents the bots you want to test against. After editing these arrays (never bothered to make a config file for this) you start RunBattles with robocode.jar in classpath and the usual Robocode-arguments (for memory and cache). For each challenger a file will be created (in the directory you started from) that contains all the information about the battles that were run. Feel free to do whatever you want with the code. --Mue
Thanks man! -- PEZ
Only one thing... My machine here can't resolve the web host of that URL. No idea why, but it was the same with the Acendant bot that was available from the same place. Maybe someone that can resolv that name can post the IP here? -- PEZ
Oops, i had no idea. I'll keep that in mind when releasing bots. IP is 141.20.20.55 --Mue
Works fine here, thanks a lot for this tool. -- ABC
I still can't get at the file. Can someone publish it elsewhere please? -- PEZ
Thats strange. But well, i've just mailed the file to you (to the pezius.com address). --Mue
Thanks! I published it on http://pezius.com/robocode/runbattles.zip if someone else has the same strange access problems as I have. -- PEZ
Mue, since you like to favour survival as long as it doesn't cost you rating points, what thoughts and insights on the issue of survivalism do you have? I notice that even if I can bring the fight CC vs Ascendant to a tie in terms of bullet damage. A always wins on survival anyway... -- PEZ
Well, i'm not sure i understand your question. I just like it when my bots survive, since the better survivor is in a certain way the stronger bot. At least that was my intuition when i started with robocode: suviving is the goal and everything else (dealing damage etc.) are just means to achieve it.
I dont remember doing anything specific to boost the survival of Ascendant. Maybe apart from using lower bullet power when having low energy. And of course Ascendant will never disable itself by firing. --Mue
OK. That was what I asked about so you seemed to understand the question. =) But I don't agree that survival is the goal. The goal is to win the match as big as you can. Survival and bullet damage are both means to that end. I have experimented some with allowing my bot to shoot itself disabled. This was when I needed bytes in Aristocles. To my surprise survival increased when I removed the code that held fire when my bot was really low on energy. I guess if your gun is accurate enough then that last bullet can sometimes save you. -- PEZ
Well i just said that i intuitively considered surviving to be the goal when i started robocoding. Right now i think surviving is nice (since there are points to get :-), but winning battles on score is more important of course. --Mue
I think your choice of distancing in previous Ascendants was cool. Very agressive bot, so hard to hit, dealing blows en masse. So cool. Like a really good boxer if you like. =) Good it seems Ascendant works best in that agressive mode. -- PEZ
Yeah, i agree completely :-). Maybe its really easier to surf the waves at a close distance (not too close of course), since there are less bullets to consider and targeting gets easier too. --Mue
I don't know. CC gets better results if it keeps a bit more distance (one of the recent non-advertised test versions has tried what I think is your distanceing scheme). Whatever distance you choose you have to trust you gun anyway. And your gun is obviously trust-worthy. -- PEZ
Congrats on improving your bot again. The air is really thin at 2060+. What did you change in 0.9.7? -- PEZ
I dont really think that this is an improvement. Version 0.9.2 also got this score but then dropped two points after i reinserted it. So i'll wait this time until at least 800 battles. And since all my tweaks to bullet powers and distance seem to fail i just reverted to version 0.9.3 but kept the optimizations from 0.9.6 (which should slightly speed up the bot without changing its behaviour). Now i'm done with bullet powers and distance and will turn to other experiments. --Mue
But the rating drifts. You'll have to compare 0.9.7 with 0.9.2's current rating to judge. A faster 0.9.3 is an improvement any way. =) -- PEZ
2078.66 after 289 battles... the bar is raised again. If it holds, all hail the new king! Even if it doesn't, congrats on a very impressive release. --David Alves
Congrats, 2085, great work. Now to see if I can find 10 points in my movement... -- ABC
Wow! Amazingly strong gun you have there. Big congrats from snowy Sweden! -- PEZ
Thank you, it feels great to be the king :-). Especially since this time its not that close (compared to last time when CC had basically the same rating). So take your time ABC, i'd like to enjoy this position some more ;-) --Mue
Don't worry, it's always much easier to talk about finding 10 points than to actually find them... Congrats again, enjoy the throne. :) -- ABC
- Yeah, i figured that out already :-) --Mue
Taking the rating drift into account, this must be the strongest bot ever. Congratulations ! -- GrubbmGait
Amazingly works, Mue, so cool i just wake up from a dream then the world changed, 208x, the 2100 is not far now. -- iiley
Thank you :-). My todo list is almost empty right now, so dont expect any improvement soon. I'll probably need some ideas first. --Mue
I knew your upgraded gun would deliver a few more points. Congratulations on raising the bar! -- PEZ
Thanks :-). I wasnt sure of the result, due to the rather sudden drop of about 10 RRGC points. Didnt know whether the rating drift was responsible for all these points, but now it looks like it was. Another good thing is that Ascendant is doing better in long battles now, still far from beating Shadow but nice nevertheless. --Mue
You should update the comment on this bots entry on the RobocodeRepository. I think I see two errors there in the very brief info provided. =) -- PEZ
Hehe, didnt know that someone reads these descriptions :-). But you are right, it was in need of an update. --Mue
VirtualGuns? I didn't realise you still had that. And I still see one error in that info. =) -- PEZ
I just reintroduced virtual guns in version 1.2, since i'm looking for a way to improve performance vs other top bots (especially Shadow and Pear which still beat Ascendant). Right now this costs 9 rating points, but i'm looking for a way to reduce this number.
Apart from that i wonder what error you see in the description on the RobocodeRepository. Something with spelling or word order? --Mue
"A quite competitive one on one megabot ..." -- PEZ
And. Amazingly you can lose 9 points and still occupy the #1 slot with a margin. -- PEZ
Ah ok, removed the 'quite' now :-). And although its still #1 i'd of course prefer to dont lose any points when using virtual guns. --Mue
Have you confirmed that your anti-surfer gun "works" against surfers when used as the only gun in the array? -- PEZ
Yep, did that. Using only the anti surfer gun Ascendant gained on average 4 percent in short battles vs CassiusClay, SilverSurfer, Shadow, RaikoMX and Shiva. It also makes long battles vs Shadow a lot closer (Ascendant gets 48%-49% instead of 40% in 1000 rounds). --Mue
- [View source↑]
- [History↑]
