[Home]GhostBuster

Robo Home | Changes | Preferences | AllPages

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

Changed: 91c91,93
I can't find the Dave's PBI discussion. I know I promised to look into if I could implement it in the ranking servlets, but I can't finde the page... What was the formula again? -- PEZ
I can't find the Dave's PBI discussion. I know I promised to look into if I could implement it in the ranking servlets, but I can't finde the page... What was the formula again? -- PEZ

The page was Requests, and the formula was (actual ranking difference - predicted ranking difference). -- Jamougha

ABC referred to the collective of RoboRumble/Participants as "the ghost". Bringing forward a succeslful bot in the rumble could therefore be thought of as it needs to be a "ghost buster". I've strated this page so we could share ideas about the activity of busting that ghost. -- PEZ

One thing I've explored a little now is the notion that WaveSurfing should make you a victim to PatternMatching guns. The surfing should leave patterns while surfing that is easy for PM guns to find. I don't think that's necessarily true. Checking the details page of some surfing and non-surfing top bots gives the story that against, say, Cigaret the surfers are more succesful than the random movers. It's relative to the success against GF targeters that GF surfing looks bad against PM guns I think. What I'm saying is that unless you find a way to surf PM as well as GF waves this is not the "ghost". Just a thought. -- PEZ

See GlowBlowAPM - the APM stands for Anti-Patter-Matchers if i remember correctly. If you could combine GF WaveSurfing with PM WaveSurfing ...--deathcon

I think I'll leave that for rozu to experiment with. I have trouble implementing PM for targeting and I can just imagine the troubles I'd run into trying to use it for movement. =) -- PEZ

The surfers are more successful against Cigaret than the random movers, but every single surfer in the rumble is a ranking booster to Cigaret. My current theory about GF vs PM: GF guns detect tendencies, PM guns detect intentions. What that means exactly or how to fight "the ghost", I don't know. -- ABC

I still don't think the ghost for a surfer are PM guns. -- PEZ

Some of those rammers show up with high PBI's and I think I have quite a few times started to work on anti-rammer ideas. But I think that's not a good idea. A reasonably simple rammer, like "goto(enemy.x, enemy.y); fireFullLead?()", will be almost impossible to get more than say 70% against. But since this rammer will end up really low ranked a 2K+ bot will be "expected" to win with more like 85% against it. It's a bit like picking down the moon, wouldn't you say?

Hmmm, so I won't work with anti-PM and anti-ramming... My brain is drained! This shouldn't be this hard to figure out. -- PEZ

At the risk of repeating myself and once again disclosing some ignorance, I still maintain that PatternMatching guns are able to trigger on decelleration in a deterministic, that it is to say WaveSurfing, movement system. I some times think that people lose sight of the fact that WaveSurfing will behave the same way every time that the specific conditions are met. There is no built in randomness to the movement. Given a deep enough buffer, that is some how "cross-sectioned" enough, your movement should be toast to a PM gun. At some point, deterministic movement will be punished by properly figured GuessFactor gun too. It is only a matter of time until he figures it out. -- jim

But it will take a host of such guns before it gives any effect on the general high ranking of the surfers. And by then we might have countered in some way. And, there is the possibility that the same conditions will very seldomly be met in a 35 round battle anyway. -- PEZ

Of cause PM work really great against wave surfers. Just look at the negative PBI of almost every wave surfer against cx.* bots. Most of them use illeys wave-pm implementation. About the rammers, it should be no problem to beat them by a big margin. Remember that the rammer has to travel really long to get to you. If you manage to shoot him out while moving away from him, until he is able to reach u ... but thats all theory. In realitiy it doesnt work 100% in my bot Shiva : against radnor.[RamRod 1]?.0 63% and against bbo.[RamboT 0]?.3 89% --deathcon

Not at all. PMs work great against wave surfers only relative to how GF guns work. CassiusClay is my best cx.* slayer I have yet created. It at least works better than all my other bots. Even though the top wavesurfers of today are not designed to dodge PM guns they still manage to fool them better than most other movements do. And when it comes to rammers. Check the details sheet of NanoDeath, which is a very mean rammer. No bot gets more than 70% against it. Yet, say SilverFist is "expected" to get 81%. I just don't think its much you can do about that. It only takes the rammer some 80 ticks to reach you (in the unlikely event that the bots start out in opposite corners of the field) and in that time frame you can hardly shoot it to death. The average time before the rammer reaches you is probably around 35 ticks. What you could argue, I guess, is that it should be possible to at least be among the top anti-rammers. But I still think they are too few to be the ghost to bust. -- PEZ

Here's a table of Cigaret's 49 fierest opponents:

enemy% scorebattlesexpected %PBI
pe.SandboxDT_3.0239.1335.23.8
jam.mini.Raiko_0.4339.6640.8-1.1
cx.Lacrimas_1.3642.1440.51.6
pez.mini.Tityus_0.9.142.31744.4-2.1
pez.micro.Aristocles_0.3.742.4648.2-5.8
wiki.SilverFist_1.043.7431.212.5
pez.rumble.CassiusClay_1.044.133311.1
jam.RaikoMX_0.2944.4930.713.7
abc.Tron_2.0244.93554.2-9.3
dft.Immortal_1.13.145.5149.1-3.6
kawigi.mini.Fhqwhgads_1.145.73855-9.2
vic.Locke_0.5.846.3146.4-0.1
jam.micro.RaikoMicro_1.4446.4744.91.5
abc.tron3.Shadow_3.0746.9131.615.2
cx.CigaretBH_1.03472249.7-2.7
wilson.Chameleon_0.9147.23250.3-3
wiki.mini.BlackDestroyer_0.9.047.91153.4-5.5
kawigi.sbf.FloodSonnet_0.9481357.5-9.5
pe.SandboxLump_1.52483452.7-4.7
kawigi.sbf.FloodMini_1.448.13252.2-4.1
ahf.r2d2.R2d2_0.8648.1949.6-1.5
ad.Quest_0.1048.2548.6-0.3
pez.mini.Pugilist_1.9.9.4.148.2634.513.7
cx.micro.Smoke_0.9548.32653.2-4.9
axeBots.SilverSurfer_2.2748.3432.815.5
jekl.mini.BlackPearl_.9149.1652-2.8
pez.mini.VertiLeach_0.4.049.1749.9-0.7
rz.Aleph_0.3349.7535.214.5
wiki.mini.Griffon_0.149.92651.7-1.8
axeBots.Okami_1.0449.91040.89.1
cx.BlestPain_1.4150.6842.68
mue.hyperion.Hyperion_0.7.350.9349.91
davidalves.net.YALT_1.350.9242.48.5
pez.clean.Swiffer_0.2.951.41151.8-0.3
kcn.unnamed.Unnamed_1.2151.6549.81.8
bvh.mini.Wodan_0.29751.8660.3-8.5
rz.GlowBlow_2.3151.94854.2-2.3
kawigi.f.FhqwhgadsMicro_1.0523555.4-3.3
pez.frankie.Frankie_0.9.6.152.1755.1-3
sgp.JollyNinja_3.5352.23667.4-15.2
jekl.Jekyl_.7052.91751.21.6
cx.mini.Nimrod_0.55534457.5-4.5
kvk.HebusLeTroll_0.4153.2860-6.7
pez.gloom.GloomyDark_0.9.253.71152.11.6
cx.Princess_1.054.13252.41.7
bvh.jmd.Jormungand_0.3554.2362-7.7
rz.GlowBlowAPM_1.054.228531.2
tad.Dalek98_0.9854.33357.2-2.9
axeBots.Musashi_2.1854.41048.95.5

As you can see we have a surfer at the top. And there are 5 surfers in the top-10. Of course, some of the story is told by the guns. But I think we can say that surfing isn't worse than anything else against an iiley-class PM gun. In fact it's among the best strategies you can choose. I think that maybe Paul's way of surfing could be extra interesting to study though.

-- PEZ

But if you look at it from the point of view of Tron, f.e., I lost a lot of points against Cigaret by implementing wavesurfing. The random moving version is in the top 10, the wavesurfing one isn't even in the top 50. Also, with my gun (PM), many times I have more problems winning against Raiko than RaikoMX, the same for Aristocles and Pugilist and others. My point is that, against a PM gun, wavesurfing is not much better than a good random movement. -- ABC

That's my point too. Even if I phrase it the other way around "it's not worse than random movement". Something that seems to be the idea in some heads. Though I think surfing has a greater potential. Random movement only takes you so far while surfing obviously can bring you farther (like DT) and this while it is quite new and unexplored. -- PEZ

And now I have changed my mind about what the ghost is about. Not that I really know what to do about it yet, but anyway. =) -- PEZ

Well, in trying to work out what the ghost is I decided to hack together a script to give me the Dave's PBIs from the rankings page, and to compare them between bots. One of the interesting things is how much crashes affect the ratings - both sgs.[DogManSPE 1]?.1 and grybgoofy.[GoofyBotCS 1]?.0 seem to crash randomly in the first round, which can result in relative differences in the PBI of over 1000. The point-yielding bots seem much the same between all of the wavesurfers - head-on targeters :-) - but the problem bots differ pretty widely. mz.[NanoDeath 2]?.56, fnc.bandit2002.Bandit2002_4.0.2, fushi.PvP1?.[PvP1 2004]?-02-16 and radnor.[RamRod 1]?.0 are some of the most common ones, but we may all be fighting our own individual ghosts. -- Jamougha

I can't find the Dave's PBI discussion. I know I promised to look into if I could implement it in the ranking servlets, but I can't finde the page... What was the formula again? -- PEZ

The page was Requests, and the formula was (actual ranking difference - predicted ranking difference). -- Jamougha


Robo Home | Changes | Preferences | AllPages
Edit text of this page | View other revisions
Last edited June 5, 2004 3:28 EST by Jamougha (diff)
Search: