|Thread title||Replies||Last modified|
|Thoughts: gun inertia||5||11:05, 12 February 2020|
For several times I see a learning gun bumping from one direction to another direction, and yet each time an aim isn't satisfied, so no bullet is fired at all for a long time.
Actually this scenario disables the gun, until it gradually learns the new movement of the opponent.
However this behavior doesn't make sense. A prediction should be consistent, at least for a short period of time (and until new information that denies previous assumptions appears), or you're basically saying your own prediction is wrong, by changing mind frequently.
So, analog to human brains, could we just introduce an inertia to gun prediction process, making it stick to previous predictions a little bit more, and as a side effect reducing noise?
E.g., introducing some rolling average, accumulating previous predictions with the newest ones.
This could be done at firing angle, but can also be done better, by doing the "accumulation" (merging previous cluster with newest cluster, and weight them properly, or adding things up with VCS bins) before the kernel density estimation.
I spot a lot of guns (as well as movements) are having weakness when lateral direction is reversed. May the inertia help this scenario?
There are several bots which oscillate (reverse lateral velocity every tic). In this situation linear guns or any other simple predictors which use only the latest velocity information will fail miserably. I learned it in a hard way.
To a degree it all depends, how many tics of information you are using in your predictors. Some are using some sort of an average for a given number of clicks as extra input for a gun. I think raico is one of examples.
Of course the real art is to chose the correct averaging formula, because putting every tic into the gun expands search (or knn) space very quickly.
A pattern matching bot might do better, but I think none of them are sufficiently high in 1on1 ratings.
You can certainly reuse the cluster from previous knn search directly, this only makes kde taking a little more time (and for VCS kde, it's almost zero overhead). But yes, the rolling rate hyperparameter is crucial and hard to tune. But remember we have genetic tuning right? ;)
This would make sense if there were a single peak probability we were aiming at, but I think movements tend to be multimodal. If we are always aiming at the highest peak, then if the peaks move up and down then a different one will keep being the highest.
Neuromancer has some protection against this, it won't fire if the current bin the gun is aimed at is less than 80% of the highest peak. This way it tries to balance between 'ideal' targeting angle, and still firing frequently enough.
Of course, it's gun is the main weakness still, so maybe this is a terrible strategy ;-)