DarkHallow/Observations

I have been running some extensive TargetingChallenge tests with DarkHallow and I have noticed that the number of GuessFactors seems to not make much of a difference beyond about 23. Admitedly this is the TC and hardly representative of actual Robocode conditions but I think it is an interesting observation. I have tested 2 seasons each (500 rounds) for all odd GuessFactors from 23 to 47. In each instance I found that there is little or no difference between any of them. A slight uptick was to be had in the move from 23 - 25 but not significant. At smaller bin counts, using bot width to determine the correct aiming factor adversly affected the hit rate. As the number of bins increased I found that using bot width had less of a adverse affect but rarely performed better than simply always assuming that an enemy had a width of 3 GuessFactors (DH's default characteristic). I did not do any testing without some sort of bin smoothing as I have always found that performs poorly. I may need to run some of those tests to see how it works out. This has all been done with DH and the TC. Take my results with a grain of salt. -- jim

This is about what I have found too. Maybe if you are fitting your bot for battles on 1000 X 1000 fields you should opt for 29 bins or something. I haven't found that assuming a bot width of 3 bins should be better than assuming 1 bin though. Not that I have tested it very much. -- PEZ

I have done some limited testing of no smoothing and I find that assuming a width of 3 bins seems to make me slightly more accurate in the short run but has no effect at all in the long run. I am not sure why this is. Maybe it helps to "smear" the data when you do not have many observations but does not hurt to have the data "smeared" when you have a lot. I find all of this interesting. My initial assumption was that the more bins I used the more finely chopped my data would be and there for the data would show more spikes. Now I am wondering if I have over segmented any of my other segments as well. -- jim

What is the difference between what you call bin smoothing and assuming the bot width is 3? Isn't the last actually a form of bin smoothing? My testing has showed that there is a relationship between the number of bins and accounting for botwidth. At one time, I had a low number of bins (I think it was 31). Then I tried a much higher number (75). My rating dropped. I reverted that, and introduced accounting for bot width. My rating dropped. Then I combined 75 bins with accounting for botwidth. Bingo! My rating improved. Note that I account for bot width dynamically, taking the distance into account. Another thing: I found that accounting for botwidth has two pitfalls which are easy to overlook. You can find these if you put your bot against sitting duck after implementing botwidth accounting. If you contantly hit it on its side and occasionally even miss completely then you have fallen for both of them :-)

1. make sure that the outermost bins, when converted back to angles, really fall within the bot's width. Initially I overlooked that the outermost bins offen only partially cover the enemy. If the middle of those bins (which you will actually use for aiming) falls outside the bot's width, then you could miss your target.
2. when you have multiple neighbouring bins with the same highest number of visits, then you should make sure that you choose the middle one of those bins. In theory it shouldn't matter, but practise show that it does, because when I remove it my rating drops.

--Vic