New Gun

Fragment of a discussion from Talk:Nene
Jump to navigation Jump to search

I've done some work on algorithms to give optimal firing power. My method itself seems quite effective for the most part except there are a few caveats:

  • Optimizing for score, optimizing for survival, and optimizing for PL yield rather different results for bullet power. The score vs survival difference is obvious, but optimizing for PL vs score is also important because the risk/reward tradeoffs are different depending on the score thus far in the match.
  • It depends on accurate numbers for hit rates. For this I've implemented hitrate calculations which factor out the impact of distance on hit rate, but being able to incorporate other aspects would be better.
  • Ideally you ought to try to model the future hitrates, by observing how the hitrates are changing over the course of the battle, however because the amount of data is limited, this can be difficult. It may be best to assume the hitrates are following a certain curve until the battle shows otherwise.
  • My modeling of outcomes at different bullet power values is easiest to implement as a monte carlo simulation of the rest of the battle. Implementing this in a bot like I have in the past causes a dilemma, because with few iterations it's accuracy is very poor, but with many iterations it starts to use a very large portion of the bot's avaliable cpu time.
  • One could approximate this modeling by curve fitting it, though that could get a bit messy because of how many input parameters influence it. In my next bot I probably will use a curve fitted approximation of my bulletpower modeling.

Part of what really has me feeling good about accurate modeling of outcomes caused by bullet power, is that my implementations have had emergent behavior which I didn't expect at all but makes sense in retrospect. For example, it turns out that if it's almost 100% certain that the opponent will win a particular round, it can be worth it to dump remaining energy into high power bullets to deny the enemy bullet damage score.

Rednaxela04:02, 1 March 2013

I'd love to see direct data of how it compares to what Diamond and DrussGT have right now. I also have well-normalized hit rate data, and even things that seem extremely safe and advantageous, like "use power 3.0 in situations where I expect a hit rate over 25% with 95% confidence" hurt me, or at least didn't help.

Voidious04:05, 1 March 2013
 

I would like to see a simplified approach to modeling that kind of behavior.

Chase04:30, 1 March 2013