CassiusClay/Real vs Virtual Waves

From Robowiki
< CassiusClay
Revision as of 08:18, 14 November 2011 by PEZ (talk | contribs) (Replies)
(diff) ← Older revision | Latest revision (diff) | Newer revision → (diff)
Jump to navigation Jump to search

A long way back in CC's history I thought it was a good idea to weigh "virtual" waves lower than "real" waves (i.e. waves that follow a bullet fired by CC). The rationale was that it makes the stats usable against bots that react to bullet fire while still being able to learn quickly against bots that do not alter their movements on enemy fire. Now, many years later, when discussing the anti surfer gunning of Dookious it struck me that I might have overlooked the fact that up-weighting the real waves might be almost the same as not using virtual waves at all. I'm not sure that this is entirely right, but there is some truth to it and it does matter (a little) in combat.

What I did to check this was to add a Virtual Gun that weighs real and virtual waves equal. (Then I managed to blur the experiment with some other stuff.) Now I have repeated the experiment without bias from other changes. I've made three versions of CC with some different gun configurations, like so:

Rumble entry Guns APS Survival Wins
2rho.02nc Virtual + real waves 82.8% 91.7% 97.0%
2rho.02nd Real waves 82.4% 91.3% 97.1%
2rho.02ne Virtual waves 82.4% 91.0% 95.6%

Granted, it's not a world of difference between the versions, but still, a .4% APS gain is a .4% APS gain. =) I'm not sure what conclusions I dare draw from this, but my gut says it might be something like so:

The fast learning gained from using virtual waves only is worth exploiting. I've noticed that this gun is often stronger against surfer than the other CC guns. But there are enough bots in the rumble that do react on bullet fire in a direct way so that a gun using real waves (or weighing real waves significantly higher than virtual waves is worth exploiting too.

Or:

Maybe it is more a question of the virtual wave gun being strong the first few rounds and then the slower learning guns can pick up from there?

Or:

... Your thoughts here. =)

(Note that all three versions have a partner gun with really fast rolling of stats. It weighs real waves > virtual waves. I might conduct some more experiments later investigating what this gun contributes or not.)

Recent waves more significant

I feel as though using virtual wave information is better than not using it. But I haven't played with weighting them differently. I weight virtual waves the same as real waves. However, one extra thing I do is to given recent waves more significance or weighting. In my own terms, I tend to think in hit ratios. And I currently use 0.75 * overall hit ratio + 0.25 * rolling hit ratio, where the ratios include both virtual and real bullets, and the rolling ratio uses the last 40 shots. I haven't played with weighting the virtual bullet ratio different from the real bullet ratio, but now you have me curious enough that I might have to play with it. -- Skotty 00:47, 14 November 2011 (UTC)

Just to make something clearer; The difference between the CC guns mentioned is how they weigh the visit counts, not the hit ratios. Nonetheless, you just pointed out something important to me, that I should consider mixing in the hit ratios of the virtual waves. (I currently only measure hits from real waves.) -- PEZ 07:18, 14 November 2011 (UTC)

Switch to real waves when they have enough data

I think there might be some benefit in using all waves until they reach a certain number, then switching to real waves only. Once there are enough real waves the virtual waves can be dropped entirely. This should capture the best of both worlds, without having to add an extra gun. --Skilgannon 06:15, 14 November 2011 (UTC)

This is worth investigating for me. I should at least try to not penalize the real waves with the potentially lower hit rates of their initial rounds. But I still think virtual waves might keep playing an important role in later rounds against surfers since they can speed up the relearning, at least in theory. I'm running a new experiment with this as I type. Let's see what it brings home. -- PEZ 07:18, 14 November 2011 (UTC)