User talk:CelerManusDei

Welcome to the wiki, and best of wishes with Durandal :) --Rednaxela 04:58, 4 August 2009 (UTC)

Hey there, welcome to the wiki! :) I downloaded Durandal, just to check it out, and it seems pretty good for a first bot (How did wave suffering treat you?). I like the color scheme!Spinnercat 05:38, 4 August 2009 (UTC)

Thanks for the welcome guys. The wave surfing was fine thanks to the wealth of info on this site. The real challenge was the dynamic tree structure I set up for GF targeting. Hopefully it was worth it. --CelerManusDei 06:33, 4 August 2009 (UTC)

Welcome! You can put anything on your user page, maybe a little biography of you. Good luck! » Nat | Talk » 12:08, 4 August 2009 (UTC)

A belated welcome! Looks like Durandal is off to a good start. Please feel free to make a page for your bot, too, your gun sounds very interesting. Best of luck. =) --Voidious 19:46, 8 August 2009 (UTC)

Thanks! I just finished Durandal's page. Still can't beat Druss though =p Oh well. --CelerManusDei 21:22, 8 August 2009 (UTC)

Welcome! I'm also very interested in the gun, although my faith in a self-weighting tree is a bit shaky, because it needs time to find good weights as well as find which angle from those segments yields the best results. Oh, yes, sorting your results details it seems DrussGT is the bot you do worst against :-p But good luck in improving your rating, it seems like you're very capable as a programmer for something like this to be your first release =) --Skilgannon 23:38, 8 August 2009 (UTC)

Very interesting. The gun sounds rather similar to the gun I whipped up a few months back, titled "StainedGlass", which I never released with any bot but also generates a tree. It's performance as a gun was respectable, but a far cry from the top guns out there. Is Durandal's gun's tree a binary tree, or does it branch into more than two segments at each step? --Rednaxela 23:46, 8 August 2009 (UTC)

The tree tries to replicate a multi-dimensional array of guess factors without having to allocate the entire array (before coding the tree, I was trying to use a seven dimension double array, but that used way too much memory) It isn't binary, and can generate a predetermined number of branches for any given "segment." Branches are not created until a certain number of visits (per branch) is reached, and each branch contains a visit count array. The longer the match goes on, the more branches are added to the tree (up to a certain point). When a branch is formed, the parent transfers its visit count array to the child so that the child can refine the data the parent gathered. As it stands, I can probably segment on 20 or so dimensions before I start running into memory issues, though I don't know if using that many dimensions is wise. I understand that too many segments means that the gun learns too slowly, but the fact that a parent passes on a copy of its visit count array to the child means that the learning time is decreased. When I finish commenting all of the code, I'll post it somewhere if anyone wants to have a look. This probably isn't the first gun to use these methods, but my real goal when coding it was to go for minimum memory usage while still maintaining efficiency. --CelerManusDei 00:02, 9 August 2009 (UTC)

Ahh, my approach in my StainedGlass gun, in constast was a binary tree, and instead of just segmenting across a single axis per layer, it would use multidimensional least-squares linear fitting to split across the most useful linear combination of axis to split across. As far as the number of dimensions is concerned, the reason it is often said to be slow when you have more dimensions, is that you'll have too few data points in each segment. With the way that you only branch upon there being a certain number of data points, that conventional "wisdom" does NOT apply whatsoever. You can put in as many dimensions as you want Celer, and the speed at which your gun will learn will solely depend on the threshold at which you have it branch --Rednaxela 00:38, 9 August 2009 (UTC)