Robocode: Neural evolution test with NEAT

[xxKiraNLxx]

I dont really see the point of this tho, It would do terribly in matches like this, could you explain why you would use this?

Regards,

KiraNL

[Milez1337]

Spazbot.

[aidanjmorgan]

Is there any chance of seeing the source code? I'm doing something similar (involving scanning and movement) and i would love to see what you've already achieved with ANJI.

[robocodeWarrior]

check out our robocode competition codewarriorswanted

[cgmiguel]

That was a nice application of NEAT.

Do you have an article or similar on it? I'd really like to take a look.

[gistanevicius]

Well, I wrote my bachelor work on this, but it is in Lithuanian, so I guess you wont read it :)

[cgmiguel]

I guess I won't have time to learn Lithuanian right now! Too bad :(

I've trying to achieve similar results using NEAT, but I'm having problems (although almost solved) in integrating it with robocode.

What version of NEAT did you use? ANJI, NEAT4J, ... ?

[gistanevicius]

ANJI. I used robocode security settings turned off, so I could use external libraries

[cgmiguel]

I'll check out the security options as soon as I can (I'm working with my own modified NEAT4J implementation).

Do you have any intentions to release the sources?

[gistanevicius]

I can give the sources, only my implementation is very messy right now. It was ok for testing, but it may need some work to run on other computer

[cgmiguel]

Thanks! I sent you a private message with more details.

[mgsmgsmgsmgs]

@gistanevicius I used NEAT in my bachelor work too. I'd have to show off some parts of it once. It was about live simulation (triangle life forms with brains evolving using NEAT).

[vestalunch]

well done.

what was your input/output, and does the yellow tank have the green tanks position as any kind of input, or is this some statiscal magic trick?

[gistanevicius]

there were two neural networks: one for the radar and one for the gun. I don't remember what exactly input was used here, but my newest best robot uses 11 input parameters (actually too much) for the radar and 5 parameters for the gun. The information provided for the nets is like distance to the enemy, angle to the enemy, enemy heading and velocity etc.

[KingHaart]

I know this is an old video now, but I've also mixed ANJI and Robocode and can confirm that for Radar, only 3 inputs (including a bias input) are needed to get a perfect lock on the enemy.

Targetting is a different matter, working on that now.