NERO Machine Learning Tournament

#1

Post by lucasmoellers » Fri, Dec 02, 2011, 5:08 PM UTC

NERO Machine Learning Tournament
code.google.com

Re: NERO Machine Learning Tournament

#2

Post by pitiburi » Fri, Dec 02, 2011, 5:29 PM UTC

You, sir, have just stolen many many hours of my life. I am very thankful, this is awesome. My wife, on the other hand, doesn't know you but she hates you already.

Re: NERO Machine Learning Tournament

#3

Post by lucasmoellers » Fri, Dec 02, 2011, 5:44 PM UTC

pitiburi wrote: post #2

You, sir, have just stolen many many hours of my life. I am very thankful, this is awesome. My wife, on the other hand, doesn't know you but she hates you already.

Yeah, it is probably a good thing that there are less than two weeks to submit an entry.

Re: NERO Machine Learning Tournament

#4

Post by tansey » Fri, Dec 02, 2011, 6:34 PM UTC

I've been working on OpenNERO this semester, along with Igor Karpov (lead dev), Adam Dziuk, Leif Johnson, and Risto Miikkulainen.

The idea of this tournament is that you are training a population of evolving agents, each controlled by a separate artificial neural network. It's effectively a strategy game where you teach your networks how to react to different inputs (flags, enemies, friends, etc.).

We'll then compete the teams against each other in a massive tournament.

Re: NERO Machine Learning Tournament

#5

Post by hesdeadjim » Fri, Dec 02, 2011, 7:17 PM UTC

Wow can't believe this is still going. I had a chance to work with the original creator, Ken Stanley, during a summer internship at a University of Texas media lab. We had a ton of fun training up teams of robots -- my favorite was making a team of cowards that would cower behind walls when they would see a turret.

This project and the original used a variant of his rtNeat neural network software:

http://nn.cs.utexas.edu/?rtNEAT

Since then he's moved onto the University of Florida and created some new variants. One that is particularly interesting to me is cgNeat, for creating evolving, procedural content for games. It was used to evolve new weapons and particle systems for a diablo-like loot system for this game:

http://gar.eecs.ucf.edu/

Re: NERO Machine Learning Tournament

#6

Post by chlee » Fri, Dec 02, 2011, 7:53 PM UTC

tansey wrote: post #4

I've been working on OpenNERO this semester, along with Igor Karpov (lead dev), Adam Dziuk, Leif Johnson, and Risto Miikkulainen. The idea of this tournament is that you are training a population of evolving agents, each controlled by a separate artificial neural network. It's effectively a strategy game where you teach your networks how to react to different inputs (flags, enemies, friends, etc.). We'll then compete…

I worked on NERO 5 years ago with many of the same people as well.In fact, we were the first batch of [then] undergrads join the project.

I am very glad to see that the project has continued and is still going strong.

Re: NERO Machine Learning Tournament

#7

Post by sukuriant » Fri, Dec 02, 2011, 8:12 PM UTC

hesdeadjim wrote: post #5

Wow can't believe this is still going. I had a chance to work with the original creator, Ken Stanley, during a summer internship at a University of Texas media lab. We had a ton of fun training up teams of robots -- my favorite was making a team of cowards that would cower behind walls when they would see a turret. This project and the original used a variant of his rtNeat neural network software: http://nn.cs.utexas…

University of Central Florida. I took a class under him. Stanley is an awesome professor and NEAT is my favorite style neural network model

Re: NERO Machine Learning Tournament

#8

Post by gujk » Sat, Dec 03, 2011, 8:14 AM UTC

Why not build a tournament around creating a useful product/analysis instead of winning a synthetic game?