Yeah yeah... but you see, I don't have a comp-sci background, so I can't implement any of your fancy linear regression neural nets (or whatever they are). I just go with brute force. I figure that if I have more buffers in my bot than you, it'll be a better bot than yours against a larger number of bots. So far this method seems to have worked =) Instead I can focus on doing what I do best, optimizing algorithms in my bot to prevent it from skipping turns when I put all these buffers in there. I took a similar line of thought with my gun with its 10 dimensions, 2 aiming techniques and 100 extracted scans. Intuitively, I don't think this method will work quite as well with an antisurfer gun, though. -- Skilgannon
Eh, well those neural nets are kind of failing to work right now. I'm not quite sure the advantages I was seeing were simply that it made the movement more confusing to learning guns, and for that I'd be better off with a flattener. I'm now switching to plain linear regression because that should give more reliable results than the silly neural mess. One note though, is that while our perferred approaches are different, they are both are ways of extracting more meaningful stats from the same raw data. Just as your numerous buffers show more information, my "guessfactor readjustment" approach also attempts to extract more meaningful stats from the raw data (by attempting to interpolate better guessfactors, instead of weighting dissimilar situations lower). In any case, your bot is on top (for now), so your approach does certainly work =) -- Rednaxela
|
![[Home]](/images/RoboWiki.png)