[Home]Lukious

Robo Home | Changes | Preferences | AllPages

Sub-pages: Version History

Bot Name

Lukious

Author

Voidious

Extends

AdvancedRobot

What's special about it?

It uses log based stats (DynamicClustering, a la TronsGun) for both movement and targeting.

Great, I want to try it. Where can I download it?

https://www.dijitari.com/void/robocode/voidious.Lukious_1.20.jar

How competitive is it?

It should hang tough with all bots in the RoboRumble, but you never know with a first release.

How does it move and/or dodge bullets?

WaveSurfing

How does it fire?

DynamicClustering

How does the melee strategy differ from one-on-one strategy?

No melee strategy.

How does it select a target to attack/avoid in melee?

It thinks they're the same target.

What does it save between rounds and matches?

Everything between rounds, nothing between matches.

Where did you get the name?

It's named after [Luke Skywalker].

Can I use your code?

Yes, it's released under the RWPCL.

What's next for your robot?

Code-wise:

Otherwise:

Does it have any WhiteWhales?

Not yet!

What other robot(s) is it based on?


Comments, questions, feedback:

Congrats on top-10 with the first version of this. I have a few questions. Since I've known about this project a while I got a bit precise that it includes DynamicClustering. Or if it is I who don't understant that concept correctly. Could you elaborate some on the DC implementation of your gun/movement? Also, I saw that the TC2K6 scores wasn't stellar. Can we expect MC2K6 scores soon? Or, even better, hook Dookious' gun on this movement and see what happens. -- PEZ

Thanks. As I understand DynamicClustering, you also use it in Ali, right? The cluster is the n most-similar scans, and that cluster is what's used to decide targeting/movement. Ali doesn't use anything like gaussian KernelDensity estimations, but if you look at those formulas Albert posted, you could say that the "Uniform" formula is what it uses to decide firing angle.

The only thing unique about my DC implementation is that I am using a 2D formula for KernelDensity estimations, with one dimension as the GuessFactor, and the other as the difference in distance from the current scan. I haven't done extensive research to show that it works better, but it seems to in small amounts of research, and it feels better to me (intuitively) to weight the scans by distance somehow. I also don't have to worry much about the number of scans in my cluster, since the distance stuff is taken into account by the KernelDensity stuff.

I have a problem with the MC2K6 where Lukious gets REALLY REALLY slow after a couple hundred rounds against the learning guns, because he's processing data on the scans every single tick. (I can and will come up with a way to cache the closest scans...) I can tell you that his most recent WSC2K6 scores were 99.13 / 97.69 / 95.59 = 97.47, and the most recent full MC was 37.87 / 36.26 / 52.06 / 29.42 for the rest of them. He still has a long way to go before the gun or movement could rival Dookious, but I do think the DC-based movement shows a lot of promise.

-- Voidious

OK. As I had imagined DynamicClustering it was about dynamically weighting the segmentation dimensions. By somehow figuring out which dimensions where best "describing" a particular opponents movement. -- PEZ

Well, I wouldn't mind hearing what ABC has to say about it, but I don't think that's a necessary component of DynamicClustering. That seems like a DynamicSegmentation? type of idea that could be part of a DC system. In my mind, it seems that with segmented stats arrays, you predefine the "clusters" with your segmentations, while in these log-based systems, the clusters are formed on the fly (er, dynamically =).) -- Voidious

Yeah, maybe. But to me it's just clustering. Using predefined clusters isn't clustering, is it? That's just VisitCountStats. That's why I thought the Dynamic part was about the weighting of the dimensions. -- PEZ

Well, you're right about "clustering", and it's just terminology, but... I still say the clusters are "dynamic" even without any dynamic weighting. Each time you make a decision, you redefine the cluster. With dynamic weighting, you also redefine how you form the cluster. DCBot uses fixed weights on the different factors of the distances, as does Chalk, and both are considered to be using DynamicClustering. -- Voidious

Also, I wonder if "dynamic" is because you form a new cluster with every evaluation, instead of just redefining the clusters every so often. But I realize I should probably just shut up and wait for ABC to clarify what he meant, since he coined the term in terms of Robocode... -- Voidious

Yes, I call it dynamic even without the dynamic weighting of the dimensions. The cluster you choose is never the same, unlike in VisitCountStats. And it sounds better :). Anyway, I'm glad more people are experimenting with it, I believe it has potential to go beyond GF guns, I just never managed to realise it. Great work on Lukious, top 10 with the first version is amazing! -- ABC

OK. Thanks for the clarification. Though now you're confusing me again with that reference to GF guns... Ali uses a DynamicClustering gun right? But to say it's not a GF gun would be quite misleading. In fact Ali's gun and CassiusClay's are very, very similar. The only thing that really separates them is the clustering part. Dynamic vs Fixed clustering if you like. They are both StatisticalTargeting guns using GuessFactors. Yeah, yeah, only terminology. But I'm sort of big on terminology. =) But factwise I agree with you. Intuitively DC should have a higher potential than fixed buffers. And I think the key to harvest a bigchunk of the potential is DynamicClusterDimensions?. Which dimensions to choose and how to weight them. Though I guess both can be expressed in weights (the ones you don't want at all you just weight to zero). -- PEZ

Sorry, GF guns as in "fixed clusters statistical guns using guess factors and stats buffers", as in the big majority of the rumble top bots, from FloodMini do Dookious. I just got used to calling that a GF gun ;). -- ABC

Big majority? I count to 5 of the top-10 using DynamicClustering. That's only a big majority in US presidential elections. =) -- PEZ

Hehe, but that is a recent trend, for a long time Shadow was very lonely at the top... :) -- ABC

Well, SilverSurfer has been there for quite a while. But I'll stop giving you a hard time now that you have realized your clustering method is now mainstream. =) -- PEZ

<deleted absurd comment> You don't even need a majority vote to win a US presidential election, but the winner usually gets the majority. But seriously, I too think that the DC system could have a higher potential than segmented stats buffers, both in learning speed and accuracy. Time will tell if I can figure out how to make it happen, though! -- Voidious

Ali's gun is not a traditional GF gun, that's for sure. Ali + traditional GF gun = CC. -- ABC

Hmm... yeah, I know how Ali's gun works, I think I have a faulty memory chip in my brain... -- Voidious

So you guys are saying that the "fixed clusters statistical gun using guess factors and stats buffers" a.k.a traditional GF-gun I just started working on, actually already is outdated? -- GrubbmGait

If you consider that the top 3 bots is the rumble use it, I wouldn't say it's outdated, no. Many other gun variants have been attempted, but until someone manages to outscore those 3 bots, it's very much state of the art actually. -- ABC

I think that what I'm saying is that it's hard to see where to make any quantum leaps in performance with traditional VisitCountStats. It's about striking that gold source of the exact right segentations and granularities. But DC-type stats seems to have the potential of getting us out of that boring state. It's yet to be proven though.

Note that I stubbornly avoid Ali's gun a non-traditional GF gun. Everyting about the GFs there actully is very traditional. It's only the storagre of scan info that's different from the "traditional" VisitCountStats path. Also note that TronsGun was around long before FloodMini so which one is more traditional anyway?

-- DonQuijote

So, for you, the most defining characteristic of a GF gun (as invented by Paul Evans and made mainstream by Kawigi) is the guess factor computation and not the statistical storage/clustering method? Not for me, a "traditional" GF gun is the combination GF+VisitCountStats, anything that deviates from that is not traditional GF targeting. -- ABC

It's because I come from Marshmallow which used several guns of which one was a bit like LaserTargeting. That is, it wasn't using GFs per se, but VisitCountStats definately. It also used Waves (indeed CustomEvent based Waves even). But it was all before I had heard of GFs or Waves. Once I understood the GF concept I started using that instead of raw bearing offsets. (Indeed I don't even use "real" GFs today even.) This main Marshmallow gun was the first gun I ever thought up when going beyond HeadOnTargeting. So it was very, very much like what you call a traditional GF gun. But it wasn't using GFs so it doesn't click with me to call it a GF gun traditional or not. I could very easily, extremely easily even, exchange the GFs in Ali for raw bearing offsets a la Marshmallow. I could do the same in CassiusClay. I guess for me the distinctive divides are:

  1. DynamicClustering vs VisitCountStats
  2. GuessFactors vs ScanLogTracing? (relink that if there's already a page for it).
I think the only not-so-clever combination here is VisitCountStats + ScanLogTracing?, but I might be wrong on that.

As for "traditional". The only thing that makes VisitCountStats more traditional would be that it was OpenSource a while before Musashi was released. Then again Axe was one for creating huge amounts of code which made the impact of Musashi being OpenSource a bit less than it might have been if Ali would have been released earlier. But Ali has been around a while now too. I think you might confuse "traditional" with not-choosen-by-ABC. =)

-- PEZ

Ok, I see where you'r coming from. But for me the defining moment was when Paul described SandboxDT's gun after the IBM Rumble ended, probably before your time even. He was using GFs, virtual bullets and VisitCountStats at that time. It was leagues ahead of anything done until then, and didn't win the competition only because of bad luck (10 round battles). The best gun before that was a simple PatternMatcher. It took a while before anyone could understand it, but the traditional GF gun (for me) was pretty much defined, only the addition of waves (Iiley was probably the first to describe them) changed since then. Another milestone bot, imho, was Cigaret, it used waves and probably some ScanLogTracing?, but I never could understand the code (mini = obfuscated, for me) and Iiley never managed to explain it well enough. Iirc, the GF gun mania started with David Alves and his Duelist series, Iiley's FloodGrapher, and then Kawigi's FloodMini.

Your right about the not-chosen-by-ABC part, tough ;). I made TronsGun in an attempt at being different, I tend to avoid the "mainstream" . :) --ABC

Cigaret uses a scan-log with "incremental" GFs. The only thing really different from Ali is that instead of picking the best N matches, like Ali does, Cigaret picks the best match (using squared euclidian distances) and uses that for backtracing what GF to fire at. I think DuelistMicro does much the same thing. Actually Paul didn't describe SandboxDT's gun back then. He described SandboxLump. But you're correct it was before my time. And FloodMini probably didn't spring out of empty air. I would think SandboxMini being open source had a lot to do with it. What FloodMini brought to me was an understanding on how simple you could do stuff that I had done in such extreme complexity in Marshmallow. The main difference probably was this wiki. With FloodMini I had both the bot code and the wiki-active author. I never really understood the code in SandboxMini, though I guess that today it would be like reading a textbook for children.

Anyhow, just because the first well-known GF guns were also using VisitCountStats there's no reason to merry the two concepts like they are the same. And just because the first well-known DynamicClustering gun used log backtracing there's no reason to merry those concepts either. I've seen people call DynamicClustering for PatternMatching just because they see the common log-back-tracing there. I work for a nomenclature/terminology where we try to identify the basic techniques and describe them in a way where people don't hesitate to combine them. I think the reason it felt like I was a genious innovator when I created Ali's gun was that what I did was parting with the established GFs == VisitCountStats misconceptions (or the DynamicClusteriung? == ScanLogTracing? if you like). If the wiki community had separated the basic concepts earlier I am sure we would been farther on our quest for better gunning today.

-- PEZ

How would you defined Toad's gun?

-- Florent

A "tree-based dynamic clustering/segmentation gf gun"? You made it, you name it :).

I agree with you PEZ, GF gun does not necessarily mean VisitCountStats. But still, if you go to the TC results page, f.e., most (if not all) bots with GF in the gun type column are "traditional GF guns", as in Waves+VisitCountStats+GFs... -- ABC

I think ToadsGun? has a nice ring to it =) About the "traditional GF" gun stuff... for me, "traditional" implies that it is something common, so "traditional GF" implies segmented stats arrays using GuessFactors. Heck, even "traditional learning gun" would probably seem like it means that kind of gun to me, too. But I like PEZ's point that it may be good to be more descriptive in our terminology, even if it were only for the sake of newcomers. -- Voidious

Yes, ABC. That's why I have started to not brand my gun "GF" in the TC listings. I've been itching to change all the other "GF" references too, but I'm waiting for a consensus. Instead of "Gun Type" I would like for the column to be labeled "Gun Tecnhiques". Then CC's entry would say "VCS, GF, VG". And Ali's would be "DC, GF". Shadow's would be "DC, GF" too?

Florent, that gun sounds really interesting. I've been wanting to create something like that for a while, but I'm totally in the dark at how to start. You have described it in more detail somewhere? If not, you should. You're breaking new ground. On promising land I would say.

-- ~~~~

Shadow's gun is not GF, It's exactly like DCBot, with slightly different dimensions and a VG-like decision for the gunHeat dimension. Shadow's movement is DC/GF wavesurfing. -- ABC

I described the segmentation tree when I was working on it Toad/SegmentationTree, and there is also the source code for it Toad/SegmentationTreeCode. If you have more questions I'll try to make a more detailed description. -- Florent

I like that idea, PEZ. What's "VCS"? Seems like FS (fixed segmentation) and DS would be two more options for listing, and listing DC/FS/GF on the movement types would be nice, too. -- Voidious

VCS -- PEZ

Reading that page I realize for how long I have banged my head against this GF is not VSC issue. =) FS seems like an odd alternative. Ali is using that. It's the clusering that's dynamic. -- PEZ

Yeah, now that I know what VCS is, that's a better descriptor. Although I would not call what Ali does "segmentation" at all. They're just attributes, and the clustering takes the place of segmentation completely. -- Voidious

Wow, that's a novel way to think of it. =) Seriously. What's in segmentation for you? For me it is the dimensions (or attributes) you use to try define the the targeting "situation". Ali and CC uses the same dimensions for that. If they are "just" attributes for Ali then they are "just" attributes in CC too. Clustering, wether predetermined or dynamic, acts on these dimensions/attributes. Right? -- PEZ

Well, to me - and I think the literal definition is this - segmenting means breaking each of those attributes up into segments. Ali records the raw value itself, instead, and uses a formula on that to dynamically decide on which scans are closest. I would say that Ali and CC both use attributes / dimensions, but only CC actually segments them. Still just a matter of terminology ;), but that's how I understand the term "segmentation". -- Voidious

Ouch, that Lukious 1.10DCape is slooooooow :-( , but it's gun is better than expected :-) -- GrubbmGait

Yeah, I knew it would be slow - sorry. Dooki's gun is the part of him that isn't slow. ;) I've got two clients running, for what it's worth, but they are surely wasting some cycles on ola.Puffin until he's stabilized. Considering how much more work has gone into polishing Dooki's gun, I think the results are quite promising! I may resurrect Lukious from post-rewrite Dookious. (A DC arms race with Simonton sounds fun. =)) -- Voidious

Also, other conversations aside, your name isn't showing up on the changes list. But agreed, Dookious' movement is impressive. --Chase-san

I'm wondering what you would get in the rumble with Raiko's gun....just for fun, and to see how far I have to go in my movement. -- Skilgannon

Question: I understand everything in your distance formula except this one. What's the general idea here? And also, is it true that your "position ratio" is min(2, WallDistance / MEA). Thanks for sharing all your secrets! :) -- Simonton

	v4 += 1D * Math.abs(Math.min(
    		Math.abs(Math.min(scan1.getGunHeat(), 1.4) - Math.min(scan2.getGunHeat(), 1.4)), 
    		Math.abs((1.4 - (Math.min(scan1.getGunHeat(), 1.4))) - Math.min(scan2.getGunHeat(), 1.4))
    		) / .7);

Wow, I barely recognize that code! With VisitCountStats, I can wait firing waves higher, but DynamicClustering kinda requires a different approach. The idea here is to compare the "ticks away from firing time" for each scan. I usually fire at 1.9, so I assume a max gun heat of 1.4. So gun heat of 1.3 would be the same distance from firing time as 0.1. There might be an off-by-1 issue there, as well. For the position ratio (just my internal name for WallDistance - should probably rename it), I cap it at 2 in the calculation because I never use higher values, but it's actually capped at 1 in the gun. I think of it more as the GuessFactor you would be at when you hit the wall. -- Voidious

One thing I didn't even consider happening is that Luki's gun + Dooki's movement would be the PremierLeague champ, but that is apparently the case, sweeping Dookious, Phoenix, and Shadow. Its only loss is to SilverSurfer, which, interestingly, is a bot Dookious handles quite easily. -- Voidious

Well, this is interesting: I'm finally rewriting Lukious to be based off of the rewritten Dookious. I found a bug where instead of trimming the scan log whenever by 1 when it went over the max size, it wipes the whole scan log instead. This might explain why it still did well in the rumble (probably reset once or twice per match), but the TC scores just weren't up to par. -- Voidious


I've gotten really lazy on posting the results of my gun work, but it's mainly because I haven't really made any serious progress. I've got a dynamic weighting system that I think works and is bug free. I've tested it vs unweighted results and seen solid proof that it works, but I've also since fixed some major bugs with it since then. The gun performs reasonably well as long as I don't compare it to Dookious, Phoenix, David Alves/PhoenixDC, or Simonton/DCResearch. =) While I have every intention of chasing those top guns, I think I should leave it for now and get to work with the movement. -- Voidious

See, the reason you sit at the top of the rumble is because you're all about the top of the rumble. I love the challenges. I love trying to be the best in the TC's. I, though, will soon focus on the MC's. Maybe then it'll be time to watch out. (Not that David isn't already giving you enough pressure) -- Simonton

I do quite enjoy the TC's, as well, but you're right, I'll always put the RoboRumble (or real battles) over the challenges - for a few reasons. One reason (by example): Dookious and CassiusClay were outscoring Ascendant in the TC's long before their guns' rumble performance were even within 10 points of Ascendant. (At the time of Tyranius, I think Dooki / CC were already the TC kings.) Another reason: I'm sweating over my DC gun here, running tons of tests, when it is already within maybe 10-15 RoboRumble points of Dookious; meanwhile, that clearly leaves him about 70 points behind in the movement department. And if the strength of DC is in learning quicker / more accurately on limited amounts of data, movement could well be where it has more potential to outshine VisitCountStats. As for the MC's, I think they have traditionally been a far less accurate indicator of good movement than the TC's are of targeting, though I think MC2K7 could be a major step up in that regard.

Still, the fact that there aren't as many rumble points in gunnery is one of the reasons I do like the TCs. It's simply fun to make a great gun, even if it won't pay off in the current RoboRumble system. Plus it keeps you and David busy... =)

-- Voidious

Hmm ... as a side note - is it more important to decay stats for movement? My hunch is that there's a lot more learning guns out there than there are learning movements. (I don't have a great idea how to do this well in a DC system yet.) -- Simonton

Yes, it's definitely more important - though I'm saying that based mostly on experience using VisitCountStats. I also am not sure of a good way to decay movement stats with DynamicClustering. -- Voidious

Only use the last N scans instead of your full 40000? --David Alves

I think that's what I do in the current Lukious movement, but I'm not entirely sure it's the best way to go about it... There really is no exact equivalent of the low RollingAverage I've found to work well in VCS. -- Voidious

I think the only thing you could do is using absolute game time as an attribute. --Krabb

The way I do it is I use a LinkedList to store my scans, always inserting them at the top. I remove them from the bottom when the log is too big. Then I just only look at the top N if I want to do Anti-surfing. (Though I haven't done that in a while, DC seems better suited to hitting random movers than surfers) --David Alves

I'd like to see a Lukious 1.20DCape ... I can't help but wonder what it will score now that your aren't completely clearing the logs every now and again. Also, it's getting boring seeing Dookious at the top of the PL the whole time =) -- Skilgannon

The 1.21 gun doesn't use GuessFactors, path retracing, or raw bearing offsets to project a firing angle. =) So that's kinda neat, though it's nothing too major. That was showing an improvement in the TargetingChallengeRM. A bigger improvement came when I ditched the current dynamic weighting system and used the weights / attributes from 1.19.13 in the /Rewrite, which were just intuitively set / tuned.

The firing angle stuff calls it a "GuessVector?" in my code, but it's also just like a 2D GuessFactor. I store the X component and the Y component (relative to the original abs bearing) of the displacement vector and divide that by bullet time; a value of (8, 0) would be like GF1, (-8, 0) would be GF-1, (0, 8) would be moving directly away from me. Once I have the N closest scans, I convert each scan's GuessVector? into a firing angle for the current scan, ignoring any that would result in the enemy going out of bounds.

I didn't expect this to be a radical improvement over GuessFactors or path retracing, but I think it's cool and I'm glad it works pretty well. In a melee implementation, it might be good to store the displacement relative to the bot's original heading as opposed to relative to its absolute bearing to me, as enemy movements are not relative to me lke they are in 1v1.

-- Voidious

So this is sort of like GuessFactor2D? The only thing you're using this for is checking the out-of-bounds, correct? -- Skilgannon

Why not just store angle and distance, instead of x and y? Wouldn't that make the out-of-bounds check simpler? Pre-release Firebird did that, but doing the bounds checking didn't seem to help me, so I dropped it. Does it yield any benefit in your case? --David Alves

Yeah, storing the end distance (well, multiplier of original distance) plus GuessFactor would give you equivalent information. I didn't think of that until afterwards or I might've done that instead. I got 87.00 over 50 seasons (up from 86.84) in the TCRM with this as the only change - probably just barely beyond the margin of error for the numbers of seasons, and since I also feel better about it this way, I'm keeping it. -- Voidious

Hang on, if you're precise-predicting the MEA, this shouldn't really do anything, should it? You should never get an out-of-bounds shot because the MEA takes that into account....or does it? Unless in your history they were *retreating*, and now they are up against a wall or something...but shouldn't wall attributes take care of this? -- Skilgannon

Right, you'd never get an out of bounds shot with precise MEA anyway. But the past shot's GF = 1 might equal .2 radians while the current GF = 1 = .02 radians. In the new system, if the enemy were up against a wall, that data would be ignored for GF=1 instead of turning a .2 radians into a .02 radians. -- Voidious


Just looking at the TargetingChallengeRM, I noticed that Lukious has the top score against FloodMini and Cigaret, both of whom react to bullet fire. This makes me believe that you are clustering on a GunHeat dimension, which I have found *does* help against the bullet dodgers, but reduces the score drastically against almost everybody else. I'd be interested on how Lukious does in the challenge (and in the rumble) without this dimension included in the gun... -- Skilgannon

Another quick suggestion, this time to speed Lukious up: how about 'caching' the locations of scans in the gun, instead of re-calculating them every scan? By this I mean, for a dimension that is weighted more heavily, multiply the location by the weight when storing it instead of multiplying the distance by the weight each and every time you retrieve the scan? Even your "distance from zero gunheat" could be kept in a linear fashion, and just that would remove 5 decisions (from the Math.min calls) for every scan you processed, approximately 100,000 per firing tick by the end of the match. Cleaning up stuff like this would allow you to aim more than once per fire, and possibly even do what you do in Dookious with waiting for last tick to line up with this tick. Although adding a tree would still be best ;-) -- Skilgannon


Robo Home | Changes | Preferences | AllPages
Edit text of this page | View other revisions
Last edited June 22, 2008 19:41 EST by Skilgannon (diff)
Search: