Robo Home | RougeDC | Changes | Preferences | AllPages

Well, as a parallel to RougeDC/TargetingLab, here's the movement lab.

RR Version MC Version Bot A Bot B Bot C Total Comment
Gamma4 MC01 96.22 98.18 97.53 97.31 1 season (500 Rounds)
MC03 99.94 98.33 95.98 98.08 1 season (500 Rounds)
MC04 100.00 98.72 96.58 98.43 1 season (500 Rounds)
MC33 99.94 99.07 96.61 98.54 1 season (500 Rounds)

RR Version MC Version HOF SPL GRG Sub 1 WAY (Sub 2) GR3 RKM Sub 3 ASC CC CHK Sub 4 Total Comments
Alpha12 99.85 84.95 87.93 90.91 65.80 72.81 76.26 74.54 29.26 37.74 37.49 34.83 66.52 31.0 seasons
Alpha13 99.70 85.87 90.58 92.05 66.35 73.17 76.61 74.89 31.28 41.07 40.58 37.64 67.73 31.0 seasons
Alpha14 99.81 87.45 91.45 92.90 67.68 72.74 78.37 75.55 28.80 39.62 38.00 35.47 67.90 45.0 seasons
Alpha15 99.70 85.91 91.53 92.38 66.31 72.88 77.88 75.38 32.18 40.19 39.11 37.16 67.81 45.0 seasons
Gamma4 MC01 99.83 85.55 92.07 92.48 68.29 72.74 79.32 76.03 30.83 38.73 37.33 35.63 68.11 34.0 seasons
MC03 99.93 85.25 90.46 91.88 65.71 71.99 78.99 75.49 25.78 37.35 33.67 32.27 66.34 47.0 seasons
MC04 99.37 86.45 90.01 91.94 65.70 72.90 77.96 75.43 26.03 35.21 32.83 31.36 66.11 50.0 seasons
MC05 99.88 85.56 91.67 92.37 66.60 72.29 79.19 75.74 27.40 37.25 32.56 32.40 66.78 29.0 seasons
MC06 99.79 86.61 92.03 92.81 67.68 72.24 78.63 75.43 26.14 37.47 34.68 32.76 67.17 30.0 seasons
MC07 99.69 86.35 91.65 92.56 68.12 72.36 79.06 75.71 26.49 37.62 35.74 33.29 67.42 50.0 seasons
MC09 99.80 85.01 88.72 91.18 67.44 72.43 77.19 74.81 26.89 38.59 36.58 34.02 66.86 50.0 seasons
MC10 99.92 86.33 91.28 92.51 66.64 72.99 78.86 75.93 27.26 37.43 34.57 33.09 67.04 27.0 seasons
MC11 99.80 85.59 92.64 92.68 68.22 71.95 78.10 75.03 25.86 37.62 33.08 32.19 67.03 28.0 seasons
MC12 99.88 87.02 92.28 93.06 68.75 72.40 78.75 75.57 26.14 37.02 34.76 32.64 67.51 28.0 seasons
MC13 27.60 40.59 34.34 34.18 4.0 seasons
MC14 28.82 39.16 38.62 35.54 6.0 seasons
MC15 29.14 40.34 38.08 35.85 20.0 seasons
MC16 29.54 39.01 36.89 35.15 20.0 seasons
MC17 29.53 42.33 38.85 36.90 90.0 seasons
MC18 28.52 41.28 36.35 35.38 14.0 seasons
MC19 28.93 38.06 37.54 34.84 25.0 seasons
MC20 29.16 40.81 37.05 35.67 30.0 seasons
MC21 28.81 41.11 38.63 36.18 30.0 seasons
MC22 29.76 39.94 36.75 35.48 24.0 seasons
MC17r 29.71 40.63 37.85 36.06 90.0 seasons
MC17rr 29.27 41.87 39.29 36.81 90.0 seasons
MC23 30.74 40.91 38.30 36.65 90.0 seasons
MC24 30.81 41.63 38.02 36.82 90.0 seasons
MC25 30.53 41.66 39.67 37.29 90.0 seasons
MC26 30.17 42.73 39.39 37.43 90.0 seasons
MC27 31.28 42.68 39.16 37.71 47.0 seasons
MC28 31.21 41.82 39.83 37.62 38.0 seasons
MC29 30.98 42.18 38.93 37.36 34.0 seasons
MC30 99.86 85.80 91.30 92.32 66.74 72.92 77.34 75.13 27.49 37.66 35.75 33.63 66.96 30.0 seasons
MC31 99.44 86.99 90.19 92.21 65.86 73.06 79.64 76.35 28.53 38.37 35.47 34.12 67.14 50.0 seasons
MC32 99.89 86.07 91.55 92.50 65.11 73.24 78.66 75.95 26.84 37.83 36.39 33.69 66.81 30.0 seasons
MC33 99.94 85.69 91.79 92.48 67.88 73.13 79.17 76.15 26.27 38.84 36.73 33.95 67.61 30.0 seasons
MC34 99.98 85.49 91.12 92.20 65.66 73.27 78.66 75.96 26.91 38.20 34.95 33.35 66.79 50.0 seasons
MC35 99.76 86.71 90.93 92.47 64.47 72.91 78.77 75.84 26.21 37.38 35.11 32.90 66.42 30.0 seasons
MC36 99.94 86.70 91.78 92.81 64.77 73.22 80.06 76.64 25.96 36.37 33.92 32.08 66.57 40.0 seasons
MC37 99.94 86.39 91.71 92.68 65.21 72.92 79.17 76.04 25.97 35.56 34.44 31.99 66.48 60.0 seasons
MC38 99.94 86.79 91.81 92.85 64.37 72.86 79.87 76.37 25.63 36.08 34.03 31.91 66.37 30.0 seasons
MC39 99.87 86.94 93.02 93.27 65.14 73.30 79.53 76.42 26.33 36.46 34.90 32.56 66.85 60.0 seasons
MC40 99.89 87.61 92.01 93.17 63.93 72.95 80.11 76.53 23.39 34.80 30.64 29.61 65.81 30.0 seasons
MC42 99.90 87.69 92.57 93.39 63.33 72.90 79.57 76.23 24.96 37.15 33.14 31.75 66.17 60.0 seasons
MC43 99.82 87.37 93.08 93.42 62.95 73.14 79.92 76.53 25.20 33.71 32.19 30.37 65.82 30.0 seasons

Version History:


Gahhh.... even though MC04 didn't hurt head-on performance according to WSC2K6, it's HOF performance took a major nosedive in MC2K7. The changes there were helpful to the Splinter and GrubbmThree performance but not much else. The surfer performance could surely be helped a great deal by a flattener, but I should first get my scores in "Sub 1" and "Sub3" up to decent levels I think... -- Rednaxela

Haha! The upcoming MC07 adds a NeuralNetwork to the mix, and results so far seem to show it as an improvement. What could it possibly offer a DC surfing? Well, The way I see it, normal DynamicClustering, is far from optimal against simple linear-targeting. The problem, is that for it to be accurate, it needs an awful lot of data samples (at least one at just about any velocity you may expect to be moving at) to produce good results, and it's inherently incapable of making guesses at gf values it has not seen before. Well, to solve that I added a simple back-propagation net. It's designed to find general trends in how the segment data tends to affect the guessfactor. Instead of just weighting results from DC lower when the situation is further from the current situation, I also use the neural net to reposition the guessfactors that the DC spits out, to be more relevant to the current situation. Wow... my bot is really tending to hybridize things all over the place. Actually, I find hybridizing far more pleasant than tuning, it's both more interesting and seems to get me results quicker. My targeting is DC/PM and my surfing is now DC/NN. I think this could likely be the first high ranking bot that uses both DC, PM, and NN learning/prediction techniques. Anyways, I'll post the MC07 results when I have 30 seasons I think. I think there's certainly room for improvement in the neural net though, such as trying out my "multiple plane regression clustering" idea to cluster multiple neural nets in order to more accurately model multi-mode targeters. -- Rednaxela

Well, it seems like my ideas to "focus" the data with a neural net or linear regression have kind of failed. They don't seem to actually do much except confuse adaptive guns somewhat. I'm thinking now that I might attempt to overlay a couple DC-trees with different segmentations. Well... or just before I let the focusing idea die, I might try overlaying it with the non-focused data at the same time and see how that does... -- Rednaxela

Apparently, my flattener really sucks! Hooray! Time to figure out what sucks so much about it... -- Rednaxela

Try weighting your flattener 50/50 with your main stats - that's what worked best for me. -- Skilgannon

Just tried it now. That seems to improve results a bit, however I think I can still safely conclude that my flattener really sucks. I do have a few ideas to test for improving it though... -- Rednaxela
Update: Whoops, that was pure flattener again still I think. Running a new test now, with 50/50 weighting, and the cluster size for the flattener increased to 20... -- Rednaxela
Update: Okay! It appears my flattener does indeed still suck just like I thought. To fix it, I'm thinking of making it so, instead of dodging specific points the DC returns, it will dodge a "hit range profile", basically what my gun uses. In other words, I'll make my flattener into something that's more of a model of my own DC gun. At very least, It'll be better at dodging myself, and I also suspect that processing DC stats this way will give results closer to VCS too. -- Rednaxela
Update: And... that doesn't seem to help... -- Rednaxela

What attributes are you clustering your flattener on? The important ones are distance (because their guns use it), acceleration, and lateral velocity. Throwing in a time-based attribute also helps. What you're trying to do is create a copy of their gun. -- Skilgannon

Currently I use distance, lateral velocity, and advancing velocity, and a very very light segment on how many waves have been fired before this one (just strong enough to prefer new data over old). I guess I might see if acceleration helps, but I'm not sure it will because FloodGrapher appears to show me as reasonably flat in that segment too I think. -- Rednaxela

I see you've increased the number of points surfed - are you weighting them by the inverse euclidean distance squared? Have you tried other methods, like inverse 'absolute value' distance, etc? Be careful of trusting FloodGrapher on surfing matters - it takes every single wave, and not just the waves with bullets, as a strong anti-surfer gun would (Ascendant, Dookious, Phoenix, etc). Try weighting advancing velocity lower, I don't think many anti-surfer guns use that one. Time-based attributes (also weighted low) prevent pattern matchers from picking up on your movement, as well as confusing the higher-legion "every single tick" guns like Shadow and Dookious's main gun. I found that multiplying the distance by how many hits ago the data was recorded was an effective way of preferring new data over old. DrussGT/ChallengeResults shows some interesting info - take a look at the lower MC2K7 section, testing out DC movement. -- Skilgannon

Well, I actually didn't have my "prefer new data" segment in MC15/16, and adding that seems to have helped significantly. I didn't need such a segment in my normal surfing because there isn't enough data for old data to flood the nearest-neighbors, but there are far more data points for a flattener of course. So still not quite there yet, but getting closer. Multiplying the distance by how many hits ago the data was recorded sounds interesting however I prefer my approach of making it a very weakly weighted segment, because when doing DC it's possible that all of the selected data points might be very old without that preference. As far as the weighting of points, I'm currently sticking with the same thing that I tuned for my normal surfing. It's a bit overcomplicated and messy my danger formula, but I'd rather change it for both the surfing and flattening at once some other time to keep it in sync and avoid code duplication. I think that next, I'll try acceleration, wall, and time-based segments. Also, I might be wrong, but I'm pretty sure FloodGrapher only uses real bullet waves, not tick waves, because Kawigi said "FloodMini 1.4 is a fine test bot with a FloodGrapher plugin (but it's not necessarily accurate, since it fires waves every scan). The preferred way it to use kawigi.tools.FloodGrapher" on the FloodGrapher page. -- Rednaxela

Let me guess, MC20 will have the acceleration attribute, but will return the advancing velocity to MC17 levels? =) -- Skilgannon

Yep! 8 seasons have run so far though and the score is only 35.8, so it's not looking like acceleration is very helpful, but I'll wait till I have at least 20 seasons before dismissing it. -- Rednaxela

Hmm, it appears that it was both the case that I didn't give MC17 enough seasons, and that there was a functional difference. Comparing the code, it looks like I did fix a bug that prevented some normalizing of the surfing/non-surfing stats so I suppose that was a PerformanceEnhancingBug... -- Rednaxela
Update: Whoops, not a PerformanceEnhancingBug in the old version, a newly inserted but in the new version. Anyways, fixed that now, and am now testing a configuration with a far quicker rolling flattener. -- Rednaxela

Hmm, looks like increased rolling speed helps against Ascendant but not the others... Of course... not much different against anything overall really. -- Rednaxela

Aha! Wall segments seem to have a measurable effect. Of course, my flattener still sucks compared to just about anything else out there... -- Rednaxela

I find it interesting that MC35, which essentially increases smoothing across GFs helps results notably against Splinter and hurts against GrubbmGrb. I think this is because Splinter fires much less predictably and thus reduced smoothing means that RougeDC is more likely to try to slip tightly between predicted bullet locations where there isn't actually a lower chance of being fired at. I think this means that my cluster size is too small against Splinter, and also that I should perhaps do some research into [automatic bandwidth selection] for kernel density estimates. Also, another reason I believe my cluster size is too small, is because I currently suspect that in MC06 when I reduced it, the improvement I saw was only because the formula I was using to weight predicted points based on difference in situations was far from strict enough. In fact, I believe that when the formula to weight based on the distance between situations is perfect, then increasing cluster size is almost always the best choice. I think that now, I will try:
1) Increase cluster size
2) Make kernel density bandwidth depend on the amount of data points returned by the kd-tree (increased smoothing may help when there are very few data points so far)
3) Research automatic bandwidth selection
-- Rednaxela

Hmm, MC42 seems to provide fairly nice results in Sub1. Still some problems though... For example I'm not sure how I'll get even a remotely decent score against Waylander without either bringing in some crazy segments or AntiPatternMatching. I also can't seem to manage to get those scores past 94 against GrubbmGrb nor can I figure out how Voidious or David Alves get such good scores against RaikoMicro. I guess I can still take pride at doing better than anyone else against GrubbmThree, haha. Also I hope I'll be able to get my flattener to a respectable level compared to the other surfers in MovementChallenge2K7/ResultsFastLearning. Anyways... enough lamenting I suppose, haha. -- Rednaxela

I think the 94+ against GRG has something to do with fighting further/closer. I remember doing one of those improved my score against both GRG and SPL, but hurt my score against the top bots. The RaikoMicro thing also puzzles me, seeing as I do better against him in the rumble, and my gun scores worse against him in the TCRM than Phoenix. It might have to do with attack/retreat angles? I'm not sure. As to getting higher scores against Waylander, time-based segments really help (DrussGT has 3 different time-based segments in the movement). They also help for confusing the anti-surfer guns because your movement isn't following the basic segmentation pattern the antisurfer guns are designed for. However, one thing to know about this version of Waylander is that it didn't have anything to stop it from matching or more importantly replaying across rounds. As such, the delta-heading sometimes gives wacky results from a round-break, causing Waylander to tend to shoot slightly more often towards GF0 than it should. Food for thought. -- Skilgannon

Hmm, interesting thoughts on those various things. As far as closer or further, I'm suspecting further is what would help against GRG and SPL, considering that the Raiko gun probably hits them better at far range than they hit you at far range. As far as Waylander goes, while time-based segments probably help a lot I'm still a bit concerned, because it seems like everything I've done to improve Sub1 performance kills Waylander performance more and more. One thought about the RaikoMicro thing, is that while you may score better against RaikoMicro in the rumble despite a worse gun against it, it's probably something to do with the way that the MC2K7 RaikoMicro is modified to fight close range. I'm probably going to get to sleep soon, but one thing that I'm trying overnight for MC44, is making it attempt to dodge the wave hitting the front of my bot, instead of the center of me like I've been doing. I'm not certain how well this will work yet, but it is looking like it might be promising at least for Sub1, considering that 4 seasons in my Sub1 score is over 94, though I'm not terribly optimistic about if that will hold after more seasons. We'll see when I wake up in the morning though. The issue with this change is that it means it stops dodging a wave as soon as it hits the front of the bot and thus it may be more likely to get hit on the side as it moves. I do have plans to fix this though, in a way that shouldn't detract from any advantages that may be brought by trying to keep the front edge of the bot safe. -- Rednaxela

Robo Home | RougeDC | Changes | Preferences | AllPages
Edit text of this page | View other revisions
Last edited July 10, 2008 7:15 EST by Rednaxela (diff)