Hi, I first went to ai-junkie.com because my question kind of relates to Mat Buckland's Minesweepers, but that site seems to be dead now :( Hopefully someone here is familiar with his minesweepers example from the Neural Networks in Plain English. Instead of trying to sweep up mines I'm trying to develop an agent that will move towards a target and stay there - using a genetic algorithm to evolve the network weights, similar to Mat's project. I know NNs can be notoriously difficult to debug. So I was just wondering if I put a wall or a boundary around the world that the agent can't move through would I be accindentally adding an extra task for my NN to learn, i.e. it has to avoid the walls and move towards the goal? In Mat's example he has a boundary-less world that wraps the minesweepers if they go off the screen in any direction. I'm quite new to NN and learning as I go, so maybe I'm confusing myself unnecessarily but this just popped into my head. Thanks, Max.

Obstacle avoidance is a bit harder. You'll probably need to add whiskers to detect obstacles and get the NN to turn. I've done this before in my thesis; it's very cool but certainly not the best approach!



You're better off using steering behaviors.


Quick reference if you still want to go the NN route:

Making Racing Games Fun With Computational Intelligence (Read some of the papers linked from the interview.)

Thanks for your reply alexjc,

Quote:
Obstacle avoidance is a bit harder. You'll probably need to add whiskers to detect obstacles and get the NN to turn. I've done this before in my thesis; it's very cool but certainly not the best approach!

So could I combine the two things - finding the goal while avoiding the walls - by using "whiskers"? I was concerned that by having a boundary the agent would be trying to learn 2 things (avoiding the wall and moving towards the goal) with one NN which may not have worked.

I've used steering behaviours (mostly from spending ages on Craig Reynolds' site) before. I know they'd do the job fine I'm just tricking around with neural networks now to get familiar with them because they come up in a college module next term.

You should check out Ken Stanley's work, NEAT. These are evolved neural networks that learn all sorts of fun behaviors. He has some interesting movies on his web site of agents learning to avoid obstacles as well as engage in combat.

-Kirk

http://www.cs.ucf.edu/~kstanley/

Thanks KirkD I'd heard of NEAT before and I'll check out that link you posted.

In relation to my original question I've decided to get the agent to find a target in a boundary less world and then with boundaries, better to start simply and take baby steps from there on!

If you read Mat's book (not Programming Game AI By Example; the other one, I forget what it's called right now), he goes more into depth with the minesweeper example with walls etc. Though, imho, the resulting minesweepers still aren't that great.

In the end, I'd stick with what alexjc says; he's always got good advice.