Lesson in Parasitic Losses
The highlighted comment was created in this revision.
I know some of you have said things about this recently, and version 11.7 of XanderCat is a good example. To be at the top, you really have to iron out all the little things to ensure there are not any score losses, even if they seem tiny. Squeeze every last point out of every feature you implement.
In version 11.7 of XanderCat, I updated my anti-mirror components. The new version was better in a lot of ways, and able to detect a wider array of mirroring tricks. However, my overall APS dropped enough to take me from 8th down to 12th place. The reason? More false positives on mirror detection (robots briefly detected as mirroring when it was really just coincidence). My new components, able to detect mirroring in a variety of new ways, amplified the losses from the false positives. Furthermore, it means I might be able to improve further than I had previously if I can completely stamp out of those false positives in a future version.
What other ways could I eliminate minor losses and see noticeable improvements in overall APS? What other ways might you be able to do the same? My next step will be to verify whether or not I'm losing any ground do to wall collisions (based on my latest stats, XanderCat runs into walls on average 43 times on every battle! But ultimately it depends on how much energy I'm losing from it, which I should know in another day).
That's pretty interesting. You can definitely squeeze some points out of rammers, but for me that was more a matter of personal pride / fun than really going for APS. My bullet power, distancing, and kernel density formulas had a surprising amount of room for improvement even after Diamond was at #2
On wall hits, I'm curious how that turns out for you. Dookious and Diamond hit walls sometimes in 1v1. In Melee, Diamond uses precise prediction to never choose a movement option that would hit the wall in the next few ticks, and the result is a really smooth and nice and beautiful Melee wall smoothing that never hits walls but really hugs them. I tried applying this same logic to my 1v1 movement to avoid all wall hits, and it was a super huge pain to get it all working right, and then I gained no points from it. So I removed all the mess and just left it at hitting some walls sometimes.
As a thought on avoiding false positives in mirror detection, I'm pretty happy with the margin of error calculation setup I use in my flattener enablement. Basically, the hit % threshold I use to enable the flattener has a margin of error added to it, so I only enable it if I have 95% confidence that the enemy's "true" hit % is over that threshold. Maybe you could do similar and be really conservative in your detection early on and gradually get more aggressive as the battle goes on and you gather more data. I'm not sure how that will hold up with multi-mode bots but it's definitely something I'd explore.
