Talk:Melee Strategy

Credits - Melee Strategy Old wiki page: MeleeStrategy Original author(s): Kawigi

Page title change (re: InterCaps)

I think this page should be "Melee Strategy" and that the MR page should be "Minimum Risk Movement". Obviously we should have redirects because people will still use them as one word, but I think we might as well be consistent. The big reason for doing it on the old wiki was that UseMod automatically made it an internal link if you made it all one word with InterCaps like that, but MediaWiki doesn't, so there's no good reason to do it. Are others cool with this change? --Voidious 20:43, 17 November 2007 (UTC)

Well I agree, it's just that it'll take some time to get used to seeing "Wave Surfing" instead of WaveSurfing :) -- Starrynte 21:03, 17 November 2007 (UTC)

Splitting this into multiple pages

Looking over this page, I feel like many parts of it could be removed and/or put into their own pages. First and foremost, the "timeline of a Melee battle" seems like it should be a subpage (like Understanding HawkOnFire) instead. The long lists of melee movements and melee targeting algorithms are each pretty long - I'd think a list with a sentence or two description would be more fitting, with the detailed description of each system going on its own page (like Minimum Risk Movement or Circular Targeting). Anyone else agree / disagree? --Voidious 00:44, 21 November 2007 (UTC)

Sounds good to me. --AaronR 01:20, 21 November 2007 (UTC)

From old wiki's MeleeStrategy page

General

Watch this space for more! -- Kawigi

Indeed very good reading. A real melee-movement (read: AntiGravityMovement or MinimumRiskMovement) goes up to #3 on my prioritylist, a one-on-one PM-gun just has to wait a little longer. -- GrubbmGait

Again thanks! This is a candidate for the "best Robowiki page ever" title. -- PEZ

You're welcome. Added a bit more content (can you tell I'm having trouble focusing on work?). -- Kawigi

Quote: 'Add any other tips that seem to work for you ' End Quote.
What other tips, are there any tips left somewhere. This is the most informative page I have seen sofar. Although most information can be found elsewhere scattered on this wiki, this is the best concentration of useful information I've seen. It feels like I have still to write my first melee-bot. You have my vote. -- GrubbmGait

Oh, there are plenty more development tests which can be applied to melee, I'm sure I haven't exhausted them all (or even remembered all of the ones I've used). Also, is there some place where you're voting for best pages? I think if this is one of the best pages, it's because (at least when I started trying to do it) there just is a lack of information about melee here. There's a few people who are really good at it, but not alot of idea-sharing and implementation sharing. In fact, before I released Coriantumr and Shiz, I think Nimrod and TheArtOfWar were the only top-10 melee bots that were even open-source (and I think most people are a little intimidated trying to read TheArtOfWar's code, simply because of it's mass and complexity). So I guess this is a tribute to the black hole and the prying I did to get information out of the top melee guys when I was newer to it. -- Kawigi

A real pleasure to read this page. -- rozu

Indeed. It's not only because melee info was so scarce before. This page is so well written and composed too. You should consider a writer's career Kawigi. =) -- PEZ

Well, I do think I write clearly (and not all of this is easy to explain!), and an extra read over the more complicated parts usually reveals a better way to explain something. While clear writing is a forte, however, conciseness isn't, so if I get to really trying to explain something (example: the old FloodHT hype page), you should expect to read something long, clear, and often full of rambling and rants. I think that both my clarity and verbosity are partly a result of being an educated native English-speaker (in "programming" circles, this isn't critical, but some people do still have trouble expressing themselves effectively in English. That's why we don't have a big clear disertation like this on what "Wave" pattern-matching is - I'm almost positive Iiley has written one in Chinese on his site, though).

Another thing I've sacrificed for the sake of clarity and organization is order - some of you read this all the way through the first time I posted it, and I've added things to it since then, not necessarily at the end. The diff functionality of the wiki helps reduce the disadvantages of this, though :-) At any rate, I think that this page has stablized a bit, and the last set of links I added will be what really makes it more in-depth. -- Kawigi

Minor Strategy Discussion

Early in the development of Ugluk I noticed that he was doing well in melee. His movement was not complicated either.. it was basically going top speed with a slight weave and bouncing of walls at 45 degrees. While he didn't favor corners or avoid the center of the map, he lived quite a while because in order to remain a favorable target his enemies would have to chase him at top speed like a ram bot, which isn't a popular movement style. It also means he would present himself as a target option to other robots as he went on his way, but with infrequent full circle radar sweeps by the enemy Ugluk had a low chance of being chosen as the most favored target at any point. While the movement has been overhauled a few times, I try to give Ugluk an opportunity to shake pursuers in melee while not actively finishing off an opponent. --Martin Alan Pedersen

How do you make a bot that's shooting at you shoot at another bot (besides moving away from the shooting bot)? --Starrynte

A lot of bots just target the closest bot, so making another bot the closest would be one way. Some bots take energy into account, so shooting at another bot close to it might make it target that one. Really, though, there's only so much you can do... (You could also just try to kill that bot. =))-- Voidious

My latest melee movement is a pretty sound minimum risk implementation. One thing I am observing is that it is pretty weak in the very beginning. If you look at a melee round by how many opponents are left at the time, there is a large portion of it spent with 9 opponents. If your movement can't handle the opening of the round, you are going to get eliminated much more than your share. Right now in testing against the 9 best author's best bots, I am getting more than my share of Last Man Standing, but I am also getting killed first twice as much as the average. I may introduce a new opening movement and then switch after the first person dies. (I've broken it down even further in the past, but with duel style movements.) -- Martin

Right now, Smash is thrashing between enemies a little too much. Does anyone know how to fix? (code is posted at Starrynte/Smash) --Starrynte

There are two good ways to prevent this that I've seen. The first is not to switch targets unless your previous target dies or is at least 10% (for example) more desireable than the current target. In practice, this usually means until someone else is 10% closer to you than your previous target. The other way (which I believe Coriantumr does, but I can't remember for sure) is to realize that it only matters if you thrash too much if you do it right before you were going to fire - so only switch targets if your gun won't cool for another 4 or 5 turns. -- Kawigi

Ok...thnx Kawigi --Starrynte

WaveSurfing as MinimumRiskMovement

Megathanks for making this page Kawigi! Good reading. I'm not sure I agree I wouldn't put the different stages into code. Some of the ChironexFleckeri versions did that even. CF also segments it's movement and targeting stats on this. Even if the latest version only separates the melee stages from the duel one. But that's mostly a codesize issue. It seems CF is using MinimumRiskMovement as you describe it. Though I think that name is too general. I would say WaveSurfing is minimum-risk too. -- PEZ

I definitely think that WaveSurfing is a simulation-involved form of MinimumRiskMovement, in fact, I've even tried (unsuccessfully, but I think I know what my mistakes were) to implement a flattening movement with minimum-risk concepts. As for not coding the different stages, I'm specifically referring to movement - I would encourage targeting based on those phases, because segmentations are all about the observation of different movement patterns in your enemy. Since robots who weren't even coded to care how big the battle is move differently with 4 robots than with 8, it's very reasonable to segment on that. As far as coding it in movement, I think that's dangerous, mostly because it can unnecessarily specialize your bot. I think I remember David Alves saying that he preferred using fluid functions to if statements if he could help it, and this is a good example - if you have a risk function that you want to emphasize some feature more when you have more bots, you can put getOthers() into the function, or if it's something that isn't relative to specific bots, you can make it not scale to getOthers() (what I mean is that if your risk function has a part that involves iterating enemies and a part that doesn't, if you don't multiply the part that doesn't by getOthers(), it will automatically have more bearing with fewer bots on the field, because the total numbers will be smaller). -- Kawigi

I totally understand what David means by fluid functions rather than if statements. I think many of my bots are extremes of that. Aristocles, Pugilist and CassiusClay all have very little of conditionals and work without special cases as much as they can. However, optimal melee behaviour between 10 and 4 bots might not represent a fluid change of state. I say "might" since I have very little experience with this. -- PEZ

I guess that I'm more of the opinion that a fundamentally sound movement implementation will act optimally (albeit differently) in these very different cases, because the goals are primarily the same. Added a few more sections. -- Kawigi

Wave/PM

I don't think there is such a thing as Wave-PM. Unless the mega-failure Resin counts. =) -- PEZ

That's because semantically, you think of any gun for which the shot bearing is calculated from a wave as statistical. Iiley didn't think that way (and I agree with him on this). His gun was very much a pattern matcher (just as much as NanoLauLectrik's gun) - It matched a single pattern. It wasn't a pattern projector - the way it actually calculated the bearing to fire at was unrelated to the pattern, it was the bearing of a wave fired at the point in time of the end of that pattern. -- Kawigi

No, I don't, sementically, think of any gun using waves for shot bearings as statistical. Resin should prove this. And Ali/BumbleBee uses a very Cigaret like gun. Very similar actually. (Though that's a coincident if there ever were one.) It's still statistical. As is Cigaret's. In theory you could take any PM gun and use GFs for projection instead of pattern-replay. It would still be a PM gun. If Cigaret's gun is a PM gun then so are all my Tityus derived guns and FloodMini's too. The only difference is that FloodMini stores the segmented data for much faster access. Look at how the latest version of ChironexFleckeri stores its GF's and you will see a gun that could be tweaked to act exactly like Cigaret's only order of magnitudes faster.All Cigaret is doing is asking "what GF did the enemy end up at last tick that looked like this one?". That's what CFis asking too. FloodMini makes it "ticks" (plural) and uses the most common answer. Where, going from FloodMini to Cigaret, does things turn PM? But hey, remove this babbling! It's now cluttering your excellent melee strategy page. =) -- PEZ

Where, going from FloodMini to Cigaret, does things turn PM? - at the point where it's storing the last 3500 ticks of arbitrary information (plus whatever has been collected this round) and comparing each consecutive group of 72 ticks with the last 72 ticks to find the closest match, it turns into a PM (actually, it looks like he compares every 6th tick back 12 times, probably for performance reasons). Note: Feel free to move this to another, more appropriate page :-) -- Kawigi

I think the important difference between PM and Statist is the sorting. Pure PM sort every pattern to a single unique pattern, and Statist sort all patterns to be a certain number of kind(That's the segments). So in PM, there is not probability or factor, but statist is, PM search the closest match but Statist search the big probability(factor). So in this case, Cigaret's gun are pure PM. FloodMini's are pure statistical. Some combine way search the certain number of closest matchs and then segment it to count the factors was both PM and Statistical(maybe that was how Shadow gun does).

Well now, friends, very excited seen that you are focus into melee, and this is a cool page, Kawigi, great down. Maybe more melee discussing will be added on. The last version of BlestPain i tried to add the Lacrimas 1v1 movement strategy to melee, so BlestPain becamed a melee bot. But it did not good. I's hard to write a good melee movement. But when i saw PEZ and some other robocode friends are try to implement WaveSurfing in melee, i am waiting and see, I am cowered, when i was impelmenting Pear, i plan to implement WaveSurfing melee movement in it, but you know Pear's 1v1 was not enough good yet. Well however, maybe i will have a try.~:] -- iiley

WaveSurfing in Melee

Yeah, I think I knew that BlestPain was trying to do WaveSurfing in melee. It's not a bad thought, really, but it has two downfalls:

• You don't quite know who's aiming at you in the first place (although I've hinted that you can make a conservative guess)
• There are too many dang bullets sometimes.

That's not to say it won't work, though.

I hope the net effect of this page is that more discussion, development and general competition will happen in melee. -- Kawigi

Yup, i don't quite know who's aiming at me in the first place, so i can't think of every one are aiming me, i assume every robot target the closest enemy, then i just surf one bot's bullets. So there is not many dang bullets, but it is not very precision but most robots are targeting ther closest enemy in rumble. However, BlestPain are not real a sufer, it's Lacrimas's dodge movement, so it did not very good. If i want to improve it, i should base on Pear, but Pear's movement can hardly turn to melee, months ago, I'm think of just dodge enemy's bullet(The Highest of Factors) by Anti-Gravity, it must be easy to implemented than sufering(Many bullet may be easy to dodge than many Waves). -- iiley

GrudgeBot

I was thinking of making a GrudgeBot, a bot that holds a grudge against the first bot it sees in a melee battle and proceeds every round there after to go after that same bot, ignoring all the others (cept for movement of course) till the bot it has a grudge against is gone. I was gonna do this for the sheer evil of it, but at the same time the bot wouldn't be much of a contender, but instead be a bot that drags down all the other bots. (Sorry i'm having one of those mornings where you just wanna make something truely evil) -- Chase-san

Melee PM

MeleePatternMatcher doesn't seem to work yet. Any ideas? (post on MeleePatternMatching page) --Starrynte

User:Xor recently brought a discussion about building a good testbed in 1v1 which had a big impact on the way I build my 1v1 tests.

Now I question the same about melee. I build like three different beds after each minor release in RR:

• 5 HOF
• 2x (1 bot I lose against, 1 DrBob, 2 bots I score less than 60% and the rest is random)

This was a very simple strategy that I use until today and that rarely helps me distinguish the difference between two adjacent versions. Now I feel like I have to step up on this matter instead of releasing a bunch of versions everyday to evaluate it against the whole population.

Do you guys have any strategy to build melee testbeds that are worth sharing?

I usually use a mix of the top nano bots (because they let me run battles faster).

My usual test bed doesn't contain a single robot more than once. A diverse test bed seems to be better. If you really want to focus on a certain task, e.g. dodging HOT bullets, I can also recommend running battles in a 800x800 battlefield with 3 HOFs. This simulates the "middle game" where the weak bots are already dead and you have to fight to get into the 1vs1 stage.

My standard test bed looks like this: 1000x1000 HOF, DustBunny, Infinity, DoctorBob, Wallaby, KomoriNinja, Lib. This collection covers 80% of all melee movement and targeting styles. Sometimes I include SittingDroid to see whether I can fire at him fast enough. Sometimes adding a rammer is interesting too. I can also recommend testing a simple bed with nano and micro bots, and afterwards testing the same bed but replacing one of them by a top melee bot to see how your bot compares to that one.

I used to test with a 4x HOF, 3x Shiz, 1x DoctorBob, 1x GrubbmOne and a top bot eg Portia. This worked well for developing initial surfing, but I am running into limitations so now I am making efforts to run battles against the bots I have a negative KNNPBI for. However that didn't give me any gain in score, the only thing that helped was better features and fixing bugs ;-)

I was thinking, in melee it is impossible to 'starve' your enemies of points by being a perfect dodger, because they can just shoot other people. The scoring isn't done on how much bullet damage they get against *you*, but against everybody, cumulatively. So getting a good dodging movement doesn't help as much in melee as it would in 1v1.

In fact, in melee the only way to improve your score is to boost your own bullet damage and survival. And what better way to do this than to move aggressively, with some sort of wavesurfing-derived movement? That way you get the dodging at close distances so you live long enough to survive, while feeding off of bots which aren't capable of dodging at close distances because they just want to run away. So you get high hitrates, which boosts your bullet damage and you get energy bonuses which keep you alive. By making sure that you are scanning them when their gunheat is cool you know that you can catch all of their waves, and when their gunheat is high you can scan around to make sure nobody is sneaking up behind you.

Thoughts?

Makes sense. It is a very simple case (the only one I have looked at ...) but it is instructive how well DoctorBob scores in the nano melee. The best defensive bots do a little better, but its aggressive strategy has held a strong third for many years.

Yep, this is a very valid strategy, at least until there are 4 bots left. In the beginning of a round a lot of bots behave like panic, don't know where to go, don't know where to shoot. That is the time to be aggressive and score your bullet damage and up your energy level. But when you're down to 4 (or 5) bots it is time to hold your horses and change the strategy. Although cornering an opponent is a good idea, the journey to the corner will probably be fatal to you. Even in higher weightclasses DoctorBob plays a good game, with a relatively high bullet damage. Its survival is also ok, but he hardly ever reaches the last 3 because his strategy fails at that point (but what do you expect from a nano).
I would like to see a 'micro-DoctorBob' with a secundary strategy for the endgame, although its gun would probably not be very effective at larger distances.

As a side-note: To some extend Gruwel is using such a strategy. Although not as agressive as DoctorBob, it certainly does not flee and (tries to) take out everybody near it. In the endgame it starts shooting peas (distance-dependent) to try to boost survival.

I think the kind of strategy you are describing has already been used very effectively by Wallaby.

Do you have any idea why DoctorBob is so successful while other aggressive nano melee bots aren't?

Many melee bots use convenience targeting (aim at the closest opponent). Because of that, the dominant strategy for over a decade has been to stay away from everyone to avoid being targeted. If no one is shooting at you, you effectively have 100% dodging and perfect survival. And survival score counts a lot in melee (each bot you outlast means extra 50 points).

Which is the highest ranked melee bot which doesn't use either Minimum risk or Anti-Gravity Movement?

But in teamrumble, then it is a completely different story. There are 2 aggressive movement teams in the top 5. If there is a single melee bot ramming against you, shoot power 3 bullets and it is toast. If there are 3 team bots ramming against you, you're toast.

I presume you, as a game theorist, know about the Prisoner's dilemma (probably more than I do). Do you think melee bots will ever be smart enough to cooperate and form temporary teams? Or would communication be too difficult?

I'm initially skeptical it would be advantageous. Against the weaker bots, you'd only be able to eek out a small amount of extra score, while against the stronger bots you're temporarily teaming up with, you have a lot of room to gain points by killing them instead. (E.g., going from 50% to 55% is a lot easier than going from 94.9% to 99.9%.)

But upon further thought, if two top bots already know they'll get about 50% against each other whether they team up or not, why not save that battle until the end and maximize the score vs everyone else by killing them all first. I don't really see it ever happening, but maybe it could be viable.

Edit: Using the terms from the Wikipedia article, I'm trying to say that I'm not sure R > P. You may have relatively little to gain by maximizing score vs the weak bots, and you have much to gain by improving score vs the other strong bot. For the weaker bot, of course it's advantageous to team up, but it may never be for the stronger bot. (I'm also presuming that any bots trying this would be the strongest bots, which I think is likely.) (And of course this is all complicated by differing performance in Melee vs 1v1, too - teaming up with Diamond and then facing him 1v1 every round might be a bad idea.)

I know about prisoners dilemma. I also know about trigger strategies and sustained collusion.

Meleerumble is not a prisoner's dilemma. It is a constant-sum game (all bots scores always sums 100%), while prisoner's dilemma is not. In constant-sum games, collusion is never sustained.

But 2 player sub-games do have prisoner's dilemma aspects. If 2 bots collude, they both increase their chances of survival against the other 8. If 2 bots shoot against each other, they both diminish their chances of survival. If one shoots but the other not, the one shooting gains an advantage (earning points and gaining energy).

Don't expect existing bots to learn collusion as most bots have no active developers looking after them. The rumble is full of hardcore defectors and hardcore cooperators, and in prisoner's dilemma, the best response against hardcore players who pick the same strategy regardless of what others are doing is to defect against them as well.

For collusion to work, even temporary, at least 2 bots need to go for it. If you upload a single trigger strategy bot, it won't work. If you upload a bunch of trigger strategy bots, like 300, then it might work.