Density Calculations

Surfers are designed to screw up highest density calculation on purpose.

You are better off trying to shoot everywhere, except where they were in the past and were hit by your bullets (the highest peak). Peaks which formed only after your last bullet hit still works though, and the easiest way to do this is using data decay. Using less neighbors has the same effect to a lesser extent.

MN (talk)16:28, 15 December 2013

Indeed, I use a pretty low K in my anti-surfer gun. Higher than 1 though.

Also, make sure you're running enough battles to get an accurate measurement. It might be more battles than you think. More than a few bad Robocode conclusions have come from not gathering enough data.

18:40, 15 December 2013

How low is your k? I've only run a few battles, but when my hitrate goes under 8% vs a bot which can't shoot and can't flatten, and normally its above 12%, it makes me think there is something wrong. Ill run a few more to make sure.

21:33, 15 December 2013

My Anti-Surfer gun uses 4 data sets. They are k=3 of last {125 | 400 | 1500 | 4000} data points. So the least-decaying one has ~5 rounds of data, and many of the selected data points will overlap.

8% vs 12% of raw hit rate over a few battles doesn't could still be just fluctuation. If it's a normalized hit rate that might be significant... I generally run a few hundred to a few thousand battles to test a change, for what it's worth.

21:53, 15 December 2013

Its normalized calculated by DrussGT. I just did 15 battles, it seems unlikely due to random fluctuation. Is your anti-surfer gun a composite of multiple other guns? How does it work?

22:10, 15 December 2013

Well, I consider it one gun. :-) But instead of a single KNN search, it does 4 of them and combines the results.

You should be able to hazard a guess at the margin of error by looking at the distribution of those 15 values. Isn't it 1.96 * standard deviation = 95% confidence interval? Edit: Er, then divided by sqrt(num battles)? I don't know, but it's not hard to find. :-)

22:20, 15 December 2013

Does it use different trees for the different searches? Isn't that assuming a normal distribution? The mean and stdev for k = 1 are 12.12 and .34, for k = 100 (at max) its 9.26 and .36. Seems significant. Haven't tried intermediate values, k = 2 seemed to decrease performance, but I didn't do a careful study.

22:55, 15 December 2013

Yes, different trees. It does assume a normal distribution, but I think that's at least in the ballpark.

I'm not sure, but if going from k=1 to k=2+ produces a significantly different result, that seems strange. I'd probably spend some time trying to suss out bugs. Like what if you do k=2 and only consider the 2nd point, so there's still only 1 point in your kernel density? Or what if you do k=1 but then feed two of that point into your kernel density?

19:54, 16 December 2013