Editing revision 54 of RougeDC/MovementLab

Robo Home | Changes | Preferences | AllPages

Editing old revision 54. Saving this page will replace the latest revision with this text.

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

----

[[WaveSurfingChallenge2K6]]
| '''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)

[[MovementChallenge2K7/ResultsFastLearning]]
| '''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

'''Version History:'''
* '''MC01:''' The same as the movement in RougeDC Gamma4. While I didn't change the movement since Alpha15, there was a score increase, probably from a bugfix I made in my generic wave code somewhere along the way.
* '''MC02:''' Trying an all new surfing algorithm that I think should handle all waves better. Still needs work, the results were not worth recording even.
* '''MC03:''' The same all new surfing algorithm as in MC02, however using the old danger formula from MC01, and works far better than MC02. It still performs worse than MC01 for the most part though, interestingly though, it gets a HOF score among the very highest there is. To me this indicates that it's better quality surfing but that the danger formula needs work to tune it to it probably...
* '''MC04:''' Some tweaks to the danger formula. Seems to help some though the Bot C performance still isn't great.
* '''MC05:''' Remove all but one tweak to the danger formula (making it pickier about how far away the DC match is). Bah... still a lot of work needed...
* '''MC06:''' Change cluster size to 10 (from 20). Seems to have a nice effect overall.
* '''MC07:''' First try with adding a neural net to the mix. I'm now suspecting that the improvements were mostly from just making it trickier for adaptive guns to understand my movement.
* '''MC08:''' Attempts to make the neural net work better. Failed attempts not worth recording...
* '''MC09:''' Replace the neural net with linear regression... doesn't seem to do much good...
* '''MC10:''' Surf both adjusted and non-adjusted values. 
* '''MC11:''' MC06 plus one interpolated value.
* '''MC12:''' MC06 plus a distance segment.
* '''MC13:''' First attempt at building a flattener. Disable all stats except the flattener.
* '''MC14:''' 50/50 balance of flattener and normal stats. Whoops... that was still pure flattener I think...
* '''MC*MC2515:''' Fixed the 50/50 balance to work right, and increase cluster size to 20 for the flattener...
* '''MC16:''' Switch the flattener to a movement profile system matching my gun.
* '''MC17:''' Based on MC15. Add a "prefer newer data segment".
* '''MC18:''' Cut advancing velocity segment in half and add an acceleration segment.
* '''MC19:''' Remove the acceleration segment, but keep the advancing velocity cut to MC18 levels.
* '''MC20:''' Revert advancing velocity level, but re-add the acceleration segment.'37.
* '''MC21:''' Based on MC17. Cut distancing motivation in half. 
* '''MC22:''' Based on MC17. Double the distancing motivation.
* '''MC17r:''' Re-test of what should have been the same as MC17. Now, did I screw something up by accident, or was there simply not enough seasons before?
* '''MC17rr:''' MC17r done right.
* '''MC23:''' Based on MC17rr. Increase rolling speed.
* '''MC24:''' Based on MC17rr. Decrease flattener cluster size back to 10.
* '''MC25:''' Based on MC17rr. Steal the forward/backward wall segments from my gun.
* '''MC26:''' Based on MC25. Add a time-since-velocity-change segment.segment.
* '''MC27:''' Based on MC26. Make the time-since-velocity-change relative to bullet-travel-time. Tweak the weighting of that segment.
* '''MC28:''' Based on MC27. Add a weak acceleration segment.
* '''MC29:''' Based on MC27. Try tweaking the time-since-velocity-change segment some more.
* '''MC30:''' Based on MC12 (no flattener). Redesign the algorithm to get danger values. Before it was based upon an 'intuitively' created piecewise formula. Now it's based on creating a probability density estimation expressed as a mixture of gaussians, and to get the danger of a location, it integrates across the botwidth. Could use some tweaking I think, but the code and algorithm seems cleaner and more theoretically correct to me now.
* '''MC31:''' Based on MC30. Change gaussian curve width to half for the distance between situation and double for the guess factor.
* '''MC32:''' Based on MC30. Change gaussian curve width to half for the distance between situation and leave it the same for the guess factor.
* '''MC33:''' Based on MC32. Use a more accurate [http://en.wikipedia.org/wiki/Error_function erf function] for finding the area under the gaussian curve. Courtesy of a nice public domain implementation of the erf function in [http://www.cstl.nist.gov/div837/837.02/epq/index.html NISTMonte].
* '''MC34:''' Based on MC33. Cut the gaussian curve width in half again for the distance between situations. Doesn't seem to help
* '''MC35:''' Based on MC33. Increase the gaussian curve width by 50% for the guessfactor difference.
* '''MC36:''' Based on MC33. Increase cluster size to 20. The overall score is lower, however I believe it to be notable that it helps against [[Splinter]] with hurting against [[GrubbmGrb]], and helps against [[RaikoMicro]]. Besides, score losses against [[Waylander]] affect the score in this challenge far too much I think
* '''MC37:''' Based on MC36. Add the forward-wall-distance segment from my flattener. It looks like additional segmentation doesn't help given how sparse the data is for non-flattener surfing.

=== Discussion ===

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 [[Rednaxela/MultiplePlaneRegressionClustering|"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 <code><nowiki>"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"</nowiki></code> on the FloodGrapher page. -- [[Rednaxela]]

* Would you believe I've almost never used MovementProfile graphing? I tried it briefly waaay back, but not since having any decent WaveSurfing movement. I dunno, it definitely seems super useful for tuning a RandomMovement, but I just am not sure it tells you much useful info about a surfer. A WaveSurfer can control its movement profile at will, so the profile is almost 100% dictated by the gun you're facing, right? Or am I missing something? -- [[Voidious]]

* The only trouble is that this assumes that the gun you are surfing is the same as the gun that is shooting at you =) So I think, once you get your precise prediction up to scratch, the next part of the challenge is making a 'gun' that can simulate *any* other gun as accurately as possible. This of course requires tweaking weights, messing with different attributes, and adding a flattener to simulate a gun with rolling averages. But yeah, I think I've graphed DrussGT once, around 0.3.10, and that was just for laughs =) -- [[Skilgannon]]

* Well, I would say graphing is pointless against normal WaveSurfing, however I don't think that's the case if it's put in "pure" flattener mode. If I disable all non-flattener stats like I did at once point, I believe MovementProfile graphing should roughly give an idea of how good the surfing movement '''actually''' is at controlling it's profile. -- [[Rednaxela]]

* Ah, yeah, you both make good points. Thanks for humoring a senile old Robocoder. =) -- [[Voidious]]

* I wonder if it's your old age that makes you keep forgetting to re-release Dookious? ;-) And I gotta ask, [[Rednaxela]], how many hours a day are you spending on Robocode?!? The pace you're going at makes me believe that you're coding full time! I can only take so many hours a day... -- [[Skilgannon]]

* Haha, well lately my laptop is running RoboResearch nearly 24/7 but my body isn't quite spending as much time. At least 90% of the time I'm fighting off the temptation to core more ideas to improve it and be patient for running a RoboResearch test to finish. I suppose it's a good thing it takes so long to run a research test, because it forces me to pace myself so I don't get burnt out so much. Also, it is summer break, I haven't found a summer job thus far, and I lack a social life... so... well... yeah... =) -- [[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 [http://biomet.oxfordjournals.org/cgi/content/abstract/79/4/771 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]]

Summary:
Consciousness check: 1 * 9 =
This change is a minor edit. Send email notification that RougeDC/MovementLab has been changed.
(Visit Preferences to set your user name.)

View other revisions
Robo Home | RougeDC | Changes | Preferences | AllPages