View source for Talk:BeepBoop/Version History

Awesome that you got these working. I tried for a long time, hoping to get some emergent behaviors like StopNGo against linear targeting without having to program it explicitly. It never worked out the way I wanted unfortunately.

For the active bullet shadowing, I guess you have some total score algorithm that you're trying to optimise when you choose how much to sacrifice a targeting % compared to a dodging %? Is it normalised in any way on an 'unbiased' surfing stats system accuracy eg. via bullet hits? Doesn't that screw up the 'unbiased' stats collection if your algorithm is using the surfing stats to affect bullet hits?

I found the future gunheat waves only to be useful against rambots, as dodging simple targeters from a distance is easy anyway. Turning it on against other bots didn't hurt, but it's pretty expensive because it means creating and estimating the danger for 3+ virtual waves every tick. It does produce some neat emergent behaviors when it works, like turning left for a few ticks so a circular gun fires one way and then switching to turning right.

The scoring algorithm basically tries to maximize probability-of-hitting / probability-of-getting-hit, although there is an extra factor taking into account bullet power because it doesn't make sense to try to block a 0.1-power bullet with a 2-power bullet. For example, probability of hitting comes just from the aim model’s probability density function. It assumes the surfing dangers are unbiased and accurate, which of course isn't true at all in practice, but still seems to work. I guess surfing data is unavoidably biased no matter what: surfers generally only get hit in areas they think are safe. I guess bullet collisions are more unbiased, so maybe it would be a good idea to give them a higher weight than bullet hit data, but I haven't tried that.

About active bullet shielding, I'm quite interested in that you sample candidate angles from the aimer's distribution.

I've been long thinking about improving anti-surfing guns by adding true entropy this way (it works well for non-adaptive guns against surfers, but no effect yet for the anti-adaptive gun). So I'm quite curious about whether sampling itself (without the scoring of creating a good bullet shadow), instead of always firing at the max works on its own against top surfers in BeepBoop.

I did try that out a while ago before I implemented active shadowing. Yes, sampling instead of taking the argmax guess factor for the anti-surfer gun slightly improved scores in battles of BeepBoop vs itself. However, I didn't include it in a released version because scores dropped against weaker surfers, so the change didn't improve APS.

And with active bullet shadowing scoring, weak surfers make BeepBoop firing at max angles since their guns aren’t strong enough, while the strongest surfers with the strongest guns makes BeepBoop firing at sampled angles, improving both targeting performance and bullet shadow.

If my assumptions are correct, we should notice improvements in hit rate in targeting, and decreased hit rate in surfing at the same time against the strongest bots.

Never imagined that a bot can achieve that high, and still have great improvements.

Are you going to briefly explain the changes in 2.0? ;)

Getting at least 70% against everyone . . . And more than 80% survival, did you get an extra batterypack or so ?

I think it's more like adding five 0.1-APS improvements rather than one big change. I added some details to BeepBoop/Version History and BeepBoop/Understanding BeepBoop‎‎!

That is very solid. Btw how can you sure that each changes grant you positive APS since it’s very small, being hard to evaluate.

I do test changes against a big subset of the roborumble bots (with many seasons to make sure results are significant)

Thanks, that sounds practical. I’m always doing a few seasons on the full population to test out small changes, which takes a few hours on newest hardware. However, it is still hard to make 0.1 APS changes significant.

Unbelievable, almost 3% above DrussGT! Congratulations!

Thank you! Now I am wondering if 95% is achievable...

Thanks! Looks like ScalarR is catching up fast :)

Looks like if there were two best bots, APS raises very fast ;) Looking forward to seeing BeepBoop reaching 95 APS!

Ouch, after returning from 3 weeks of holiday, I find the topbots making jumps like it was 2004 . . . An awesome achievement of both of you to carve such a gap between you and the already incredible DrussGT.