Archived talk:Engineer 20061103

Sub-pages:

Version History - Archived Talk 20061103

This is an archive of past discussions. Do not edit the contents of this page. If you wish to start a new discussion or revive an old one, please do so on the current talk page.

Is this the first neural-network bot to break 2K? I don't know enough about self organizing maps to know how big a component the NN is in that. In any case, looking pretty good - it's nice to see a lot of experimental stuff in the top 20 these days. -- Voidious

Thanks Void, I needed to take a break from PowerHouse far a little while; tuning its gun segments was starting to become really tedious. Engineer is using my hack implementation of a Kohonen network, which i'm fairly certain qualifies as a neural net by its own right. Kohonen nets are (as I understand them) an effective way to automatically cluster input vectors as they come and associate each cluster with an output vector. I just did a wiki search for "Kohonen" and came up with a brief discussion about them on the NeuralTargeting page. I agree it's cool to see a lot of experimental bots entering the rumble, eventually maybe someone will unseat traditional array based structures as the king. --wcsv

Cool! When I released ScruchiPu, I was convinced NN had a great potential, but I never found time to push the idea futher. Nice to see this method can compete with other top ones. Also I think it is the first bot to use NN for movement (at least, in an efective way). Congratulations. -- Albert

Hey, thanks Albert! I actually kind of surprised myself with how well Engineer is doing; I initially made it just because I thought NNs were interesting and wanted to experiment with one. Now it seems to have become a rather serious bot. I had no idea that this may be the first bot to use a NN for movement, that's kinda cool too. What type of NN is ScruchiPu using? I looked around on here but couldn't find out. --wcsv

It uses a backpropagation NN, feeded with speed and turn angle, to predict the next speed and turn angle. Then iterates to predict the future enemy bot position and fire to it. Movement is like MicroAspid. -- Albert

I guess I broke something in 0.2.8. It's kind of irritating to see it drop 29 points in rating and significantly improve its PL score at the same time. --wcsv

Awesome stuff, a very tough bot with a very unique approach. Do you plan on developing Engineer both or working more on PowerHouse? --David Alves

Thanks David, I'll probably get back to working on PowerHouse soon in addition to continuing work on Engineer. Engineer's gun is still relatively weak in my estimation, and it's bugging me because I know that it could be much stronger. Congrats on getting Phoenix into the 2050+ range, It too is a very strong bot. --wcsv

Wow, 2041, awesome. Nice work. Do you think Engineer could / will permanently surpass PowerHouse in rating sometime? -- Voidious

Awesome bot, I like the idea of using NN as targeting and movement, the way you did it however seems to be alot more effective, did you need any sort of wave class for Engineer or did you find a way to do it without? -- Chase-san

Thanks, Chase-San. Engineer does use waves to gather it's stats; Each wave stores a "feature vector" that describes the enemy at the time of firing, then this feature vector is fed into the NN along with the GF that they are at when the wave hits.

@Voidious: Thanks man, somehow i missed your last post. I do think Engineer can surpass PowerHouse, the only question is if i'll have time to do make him do it.--wcsv

Whoah, how long as Engineer been ranked so highly in PremierLeague? I guess I don't check that ranking often enough. Number 5 is fricken sweet, man... -- Voidious

Thanks, It's been hovering around in the top 10 in the PL for a while. As best I can recall, #5 is the best so far. Unfortunately, I can't seem to squeeze any more RR points out of it though... --wcsv