[Home]Chase-san/Prototype

Robo Home | Chase-san | Changes | Preferences | AllPages

Kohonen Map Prototype

TargetingChallenge2K7/Results
Name CC RMX SHA WS WOE SurfDM FT GG RMC WLO No SurfTotal Comment
Prototype 021 66.11 75.78 72.60 76.57 76.97 73.61 87.44 78.68 81.68 79.83 78.79 81.28 77.44 2 Seasons
Prototype 016 62.21 79.77 70.07 78.50 76.60 73.43 88.26 77.32 81.83 78.07 78.43 80.78 77.10 After fixing that bug, 2 Seasons
Prototype 019 63.08 79.60 71.02 78.00 75.60 73.46 88.72 75.16 82.91 78.68 77.48 80.59 77.02 Smaller starting neighborhood, 2 Seasons
Prototype 00563.5279.6965.9680.8380.7774.1588.7376.9175.7779.4778.1779.8176.98 Rebuilt
Shadow 3.6671.7474.2168.7575.0779.4773.8586.4378.0784.5679.1570.1779.6876.76
Prototype 01462.5380.4365.8179.7378.6773.4488.1676.1973.6480.1374.9678.6276.03 .5% again, other mods
Prototype 020 61.66 75.95 72.12 76.10 75.17 72.20 88.82 77.01 79.32 77.18 75.35 79.54 75.87 Tried extending the wallDistance functions, 2 Seasons
Prototype 023 61.35 74.68 71.73 75.89 73.05 71.34 87.55 75.88 78.39 79.95 78.03 79.96 75.65 2 Seasons
Prototype 01155.7780.3362.0677.5180.7071.2788.6577.5073.3081.7075.3879.3175.29 2 Seasons, 1.4% VG Margin
Prototype 00464.1281.2464.3177.5479.8273.4188.0274.5469.2876.7476.5777.0375.22 Tuned Virtual Gun Array
Prototype 00851.5080.1266.4566.0270.4266.9088.3778.4284.8181.4177.7182.1474.52 What I get for cheating
Prototype 00361.1277.7665.6078.2377.0271.9584.3074.1071.8475.5372.4575.6473.80 First attempt at virtual gun, 2 Seasons
Prototype 002 62.88 81.37 68.05 80.60 80.21 74.6277.0469.7972.0267.1368.7370.9472.78 Second stab at anti-surfer
Prototype 015G0 61.32 81.48 64.96 78.15 80.17 73.22 77.55 70.19 71.17 69.20 68.36 71.29 72.25 Adaptive Gun Only, 2 Seasons
Prototype 00957.6982.1560.9661.0155.1463.3990.4179.1574.4679.9376.8780.1671.78 5 Seasons
Prototype 01359.8081.9761.3458.5154.8663.3088.8077.4976.5779.1978.2380.0671.68 2 Seasons, about .5% VG Margin
Prototype 00163.0177.5866.0777.3578.5372.5177.9467.3269.4668.4266.9670.0271.27 First stab at anti-surfer
Prototype 01054.7973.8061.0357.6652.3059.9188.0278.2170.7980.5775.3978.5969.25 2 Seasons, Full NN
Prototype 015G1 50.10 57.76 62.72 56.95 56.43 56.79 88.23 77.24 75.18 78.87 75.83 79.07 67.93 Standard Gun Only, 2 Seasons
Prototype 015G2 48.03 52.94 60.80 54.63 52.54 53.79 88.01 79.31 73.50 80.88 75.55 79.45 66.62 UNetwork Gun Only, 2 Seasons
Prototype 000 30.45 37.57 44.36 31.48 30.41 34.85 87.66 77.32 70.99 81.29 76.16 78.68 56.77 2 Seasons
Prototype KT0 0.22 0.90 0.38 0.61 0.67 0.56 87.26 74.03 65.94 80.98 76.52 76.95 38.75 2 Seasons

TargetingChallenge2K6/Results
Name BFly CC Chk Cig Cya DM FM Grb RMB Tig Total Comment
Shadow 3.66 99.85 72.26 96.16 85.04 74.01 95.50 90.77 90.56 89.67 86.67 88.05
Prototype 021 98.80 67.23 88.28 80.10 77.99 92.00 87.48 89.55 87.75 89.89 85.91 2 Seasons
Prototype 016 99.31 65.83 88.22 78.57 76.69 92.92 85.59 87.95 87.96 91.20 85.42 2 Seasons
Prototype 019 99.12 64.69 87.15 78.90 76.82 92.93 87.66 87.84 88.58 88.52 85.22 Smaller starting neighborhood, 2 Seasons
Prototype 020 98.59 64.33 87.40 75.20 73.47 91.81 86.93 89.91 87.02 90.15 84.48 Tried extending the wallDistance
Prototype 023 97.46 63.52 82.50 76.68 74.75 91.42 84.04 88.36 87.03 87.33 83.31 2 Seasons
Prototype 014 92.93 65.38 85.94 72.34 72.21 90.60 84.60 83.74 84.46 92.67 82.49 Not as good as I would like -.-functions, 2 Seasons
Prototype 015G0 93.29 64.97 85.81 67.85 73.22 79.06 73.57 78.88 76.89 92.37 78.59 Adaptive Gun Only, 2 Seasons
Prototype 015G1 92.41 51.00 66.75 73.49 60.63 89.01 84.41 83.63 82.68 73.61 75.76 Standard Gun Only, 2 Seasons
Prototype 015G2 91.40 49.88 62.99 71.66 57.68 90.28 85.30 84.79 84.96 71.88 75.08 UNetwork Gun Only, 2 Seasons
Prototype KT0 79.47 0.24 1.81 68.29 3.52 89.96 77.17 80.33 80.80 3.20 48.48 2 Seasons

TargetingChallenge/Results
NameDTAspTAOWSparCigTronFhqwHTTCYngwFunkTotalComment
Prototype 007 77.30 95.74 99.61 99.19 84.98 90.50 98.19 88.22 98.92 91.68 92.43 NT/GF
Shadow 3.06 80.92 91.41 99.99 98.93 84.82 91.44 96.05 89.18 97.75 89.37 91.99
Prototype 016 80.01 89.70 99.63 96.74 78.31 84.86 93.10 88.97 96.54 85.57 89.34 After fixing that error, 2 Seasons
Prototype 019 78.80 89.43 100.00 96.48 79.76 85.64 93.07 86.28 96.95 86.41 89.28 Smaller starting neighborhood, 2 Seasons
Prototype 021 77.11 90.39 99.96 96.36 79.83 85.97 92.02 86.98 96.64 85.96 89.12 2 Seasons
Prototype 020 78.25 89.21 99.80 96.19 75.62 84.72 90.57 87.30 96.37 83.97 88.20 Tried extending the wallDistance functions, 2 Seasons
Prototype 023 74.38 85.83 99.60 95.45 76.84 83.62 92.86 88.22 96.21 84.83 87.78 2 Seasons
Prototype 005 74.64 86.63 99.53 91.84 73.49 84.30 88.22 85.89 96.85 81.75 86.31
Prototype 015G2 69.54 86.09 99.24 90.06 72.23 82.64 89.13 88.41 96.16 83.27 85.68 UNetwork Gun Only, 2 Seasons
Prototype 015G1 70.60 86.22 99.41 90.98 72.85 83.10 87.51 86.74 95.63 82.58 85.56 Standard Gun Only, 2 Seasons
Prototype 014 71.90 83.55 99.13 91.68 70.62 82.12 88.16 87.63 95.22 85.56 85.56
Prototype KT0 63.80 84.23 99.80 89.05 68.19 79.40 86.65 86.86 96.73 79.26 83.40 2 Seasons
Prototype 015G0 63.87 77.79 99.08 88.23 68.35 75.60 81.67 74.15 92.33 80.37 80.14 Adaptive Gun Only, 2 Seasons

Targeting Challenge Average
NameTCTC2K6TC2K7TotalComment
Prototype 021 89.1285.9177.4484.16
Prototype 016 89.3485.4277.1083.95 Fixed Error
Prototype 019 89.2885.2277.0283.84
Prototype 020 88.2084.4875.8782.85
Prototype 023 87.7883.3175.6582.25
Prototype 014 85.5682.4976.0381.36
Prototype 015G0 80.1478.5972.2576.99 Adaptive Gun Only
Prototype 015G1 85.5675.7667.9376.42 Standard Gun Only
Prototype 015G2 85.6875.0866.6275.79 UNetwork Gun Only


Goals

Have one version of the gun rank higher then shadows in the TC, TC2K6 and the TC2K7.


Compaints, Comments, Concerns, and Phobia's

Thats... way.. awesome, wow. *Checks to make sure bullet power is actually always at three* It is! OMG, thats totally awesome, and its still only a 3 dimensional lattice. --Chase-san

Just curious - what attributes are you feeding the network, and are the outputs one per GuessFactor bin or something else? If 3D means what I think it means (one input, one hidden, and one output layer?), I think it's pretty common to only have one or two hidden layers, anyway. Nice work so far, in any case. -- Voidious

Actually no, your incorrect there my good fellow, it only has one input and one output layer. It just how the map works, the map as it we're uses dimensions like a multi-dimensional array, as so it can have more neighbors, the more neighbors a node has, well i'm not sure, but I think it updates all the data better, I do know however it gets a better aim. Currently it has a fixed map of only 10x10x10=1000 input mapping nodes. The main reason I don't put more as it just takes so much power to do the math. As when it did a wave every tick, it didn't skip rounds, but it was a slowbot, so I think I have some leadway yet on dimensional expansion.

As for attributes, every one I can, and output is just a set of x bins per node. I have Simonton to thank for that idea.--Chase-san

That was a serious improvement! (Against the surfers) -- Simonton

Actually scratch that, its skipping turns like nobody's business! If I eliminate that, it should preform a great deal better! (and totally trash all those bots) --Chase-san

So, what was the change from 0 to 1? -- Simonton

OHH, that, well, first of all I added rolling averages, then I added a few dynamic tweaks that allows the size of the variance (size of area effected by updates), be reduced with time, or rather in this case, number of updates that had happened. A few speed optimizations, nothing overly massive. Now 001 to 002 will be interesting. --Chase-san

What was that again about lossy information there Voidious? this particular version skips less rounds then Dookious does on average (as well it should!). --Chase-san

My Virtual Gun implimentation needs a little work, how to decide to switch guns so its always right (right now only if one is .01 higher then the other does it switch) and it decays at the start of each battle (its halved). How my network works I only had to add another output vector. If it turns out I need another map set to get a better gun, all I will need is another set of weights in each node. --Chase-san

Whats funny is that, my new version of Seraphim loses to the 400R version (uses Raiko's movment and simple energy management), and its like up there in the rankings, as does Virus, Tigger and just bearly GresSuffurd. So I now know the anti-surfer gun in this rocks, its just the non-sufer part I need to better. --Chase-san

Wowie, I am currently refactoring Prototype (ergo seperating it from Raiko), I have only found 1 error so far with the node system, with the node mapping, its not something that is so much wrong, as a matter of style, but its fixed to work the way I would perfer it (it was a cube with an indent, now its just a cubic network). also I need to backtrace my gf calculation and find where its dieing. --Chase-san

Between the two Seraphims, it looks strong! Keep this up and you'll give Engineer a run for the top NeuralNetwork spot =). I take it 'S' is because movement is from Seraphim? -- Skilgannon

Very Perceptive, yes indeed S does stand for Seraphim. Engineer is awesome, and I also might add closed-source, so beating it is on my to-do list. Seraphim's movement is buggy, but dispite this it is still very good. I am very comfortable with making a completely new robot around this gun as it is something new and exciting (instead of just basically remaking the same bot over and over again... ergo Genesis and Caffiene). --Chase-san

Its been hard for this gun but I have been trying to hit the non-adaptive movers better, lets hope I come up with something better soon. --Chase-san

Do you realize that this gun is scoring better than Shadow against both AM and non-AM in the TC2K7? That's serious firepower. Now you just need to hook it up to a strong surfing implementation, and you should have a 2040 - 2050 bot. I think this gun is around as strong as Engineer's. Good work! -- Skilgannon

Why stop at beating Shadow! If I work on the non-AM bots some I might be able to push it above Pear! That and the non-AM movement isn't actually better then Shadows I don't think, just in this challenge. --Chase-san

Oh nvm, wrong TC, I was looking at 2K6, 86 on that is high. --Chase-san

How are you differing the GF arrays between the AM & non-AM tuned outputs? -- Simonton

Umm, just really the smoothing is all. --Chase-san

I stole Simontons supercomputer, did 5 of 2K7 seasons in 20 minutes. --Chase-san

I must of screwed something up, its still coming out horrible after I removed the traditional GF gun, when I find out what I may revert and fix it. --Chase-san

I want to see 007 run the TC2K7. It's TC results look pretty sweet. -- Skilgannon -- 007 -- "Bond. James Bond" =)

012 ranked about 72 overall in the 2K7, and the 011 ranked well about 75 overall, the main difference you ask? the virtualgun switching margin, 011 had it switch only if the new hit rate was .014 (1.4%) better then the one it was currently on, where as 012 just went always with the best! 005 actualy has a margin of.. you might of guessed it! 0.005, or .5% so i'll try around in that area, I like 5.43656366% i'll try that (2e) --Chase-san

Seems that percent is a bad idea, 005 must of had other factors aswell. --Chase-san

Just noticed that the RoboResearch's TC2K6 isn't the full 2k6, just the random bots, otherwise I would have stats to show you all. It will only take a few extra minutes to compute the new results. :) --Chase-san

Since I can now do all this testing, and quickly, might as well put it to good use. --Chase-san

If I were you I'd revert back to 005 (or even better, 007). These small bugs that creep in are more work to find than redoing part of the code. If you wanted to, you could diff the code to see where the bug might be, but for me it's usually a += instead of a +, or somesuch. Much easier to revert than to go through all the code. -- Skilgannon

I only ran the 005 for 1 season so its probably just at the top of the ranking, and 007 got about a 62 in the 2K7, I just never posted it, however my attempt at fixing that is seen in 008. I did run a diff, and reset 014 more or less exactly to what 005 was, cept I fixed a few bugs (yes, it had preformance enhancing bugs!). So all I really have to do now it to tune the gun alot. --Chase-san

Hey Chase, how about trying a combination of Prototype and DrussGT? My movement, your gun? -- Skilgannon

Let me try to reinsert the gun from 007 that does really good vs non-adaptive, perhaps remove either the standard or the UNetwork gun. But otherwise sure thing! --Chase-san

ProtoDruss? has been born, using Druss' movement and Prototypes gun, I haven't integrated the good non-adaptive hitting gun, so it will probably in all realism rank lower then DrussGT. I just stabbed DrussGT and made it extend ProtoAcc?, and fixed the errors (ergo, all the things we're just missing bot.setTurn.. etc really), and removing the base loop, and sticking everything in that base loop into the correct place in its structure. I also removed all the gun realted stuff, well the stuff that actually fires. I also added some good power management to the gun. --Chase-san

I found a massive error in my code, and yes, it was integer division, may integer division burn in all that is damnable!! It had to do with the vital shrinking of the neighborhood over time, and well, it was 3i/10i basically, and well, that equals 0, so it was always adding 0 to the neighborhood variance! Which means the neighborhood never shrunk and got more specific! --Chase-san

The summary says it all, I put Prototype version 014 on ProtoDruss? when I threw it togeather, so it doesn't have the much more awesome scores of 016 to boost it up, I am currently testing 017 which is testing the neighborhood's decay rate, which I extended. --Chase-san

I ran both 017 and 018, 17 decreased the decay, and 018 increased it, both brought down the score, what are the chances I had it right the first time around? Now its time to fiddle with the starting variance. :) --Chase-san

Mmm, I love the smell of supercomputer in the morning. 7:30 tc2k7, 5:45 tc2k6, and a humble 4:20 tc. for 2 seasons each of course. I'll just say this is my standard test battery. So far I have run 610 battles (that I have listed) and thus 305000 rounds. --Chase-san

And then to think that I usually ran one 35-round battle against my previous version and one 35-round battle against an older WaveSerpent . . . Do remember that the rumble is about 35-round battles. But hey, I have not improved the last versions (or months) so who am I to say something about your testmethods. -- GrubbmGait

Good to see your gun has improved. My movement has improved too (I'll upload the results to the mc2k7 page when it's finished running), but so far it looks like the top bots take the biggest knock from the movement changes. And it's not a flattener =). I just fixed the time-since-dir-change segment to work off of bullet flight time, instead of a hardcoded '30'. Another few tweaks, and DrussGT 0.2 will be released, and ProtoDruss? may need an update as well =). -- Skilgannon

That sould be easy enough, it took me about 40 seconds to convert the last version (yay for a well developed structure). :) --Chase-san

KT0 was an attempt at something Simonton put me onto, it uses a set number of nodes like my NN, and updates them like them aswell, but instead of a set lattice of nodes, and thus always the same, it instead determines the ones with the nearest weights to be nearer to it. Thus I dub it Neural Clustering. But this particular build seems hopeless against the adaptive surfers. --Chase-san

I am currently running the amtchs vs 023, 022 wasn't very promising, however I am testing a core kernal calculation, this shouldn't be required given the nater of a Kohonen map, and actually should lower the score by a tad, but I decided what the heck. --Chase-san

Hey Chase-san, I have a question about that big set of computers that you had working for you. Did you have them all hooked up using RoboLeague's DivisionServer? capability? I can't seem to get it to work on my machine. --David Alves


Robo Home | Chase-san | Changes | Preferences | AllPages
Edit text of this page | View other revisions
Last edited September 22, 2007 14:56 EST by ctb-cache4-vif0.saix.net (diff)
Search: