Author Topic: Neural Net MB?  (Read 2888 times)

Offline jknilinux

  • Bot Destroyer
  • ***
  • Posts: 468
    • View Profile
Neural Net MB?
« on: October 24, 2008, 11:58:13 AM »
Has anyone tried to make a MB with an Artificial Neural Net where the individual bots are the neurons in the network?
Even though it might seem unnecessarily complex at first, I think it has a few evolutionary advantages over trying to code the entire network into a single bot. For example:

1: They can easily build an ANN of size n.

2: So long as a bot can only move by using ties (maybe if the regular move commands were disabled (disabled as in killing anything that uses them ), so movement could only be through changing tie length etc...), then they'll have to remain as a MB in order to survive. And if the only independent actions they can make are 1: being a neuron, 2: maybe whatever code is required for BP-learning and 3: shoot to death any bots in the net that shoot at others or that doesn't respond correctly to NN action potentials (eliminates cannibot & senescent cells), then they'll have to climb over a steep evolutionary hill to "evolve" back into non-ANN bots.

Anyway, even if you've failed in your implementation of something remotely like this, I'd like to see your code. I'm thinking of making this my long-term DB project, so suggestions on it are welcome.

Thanks!
« Last Edit: October 24, 2008, 10:37:57 PM by jknilinux »

Offline Moonfisher

  • Bot Overlord
  • ****
  • Posts: 592
    • View Profile
Neural Net MB?
« Reply #1 on: October 25, 2008, 08:55:38 AM »
There's a lot of issues when using ties for this :
- You can only manipulate one tie per cycle, you can't change to a new tie mid cycle.
- The individual bots will have a hard time benefitting from the network, if a bot is just aneuron it will never see the end result.
So you would have to select a certain behavior that the network will manage, like movement and action for a certain group of bots, but the idividual parts of the network should also have some basic defences and such.
- You can't use BP in such a network since your weights are all stored in different robots... or atleast it'll be very complex to excute parts of BP in each bot.

The best sugestion I can think of would be a fungus like network that forms a shell on the outer rim, but even then it will eb realy hard to manage things like shell and defences through a network... I would probably limit the network to managing some overall values and send them to all the bots or start processes for growing, moving, making shell, energy sharing and stuff like that...
Then all bots have some basic code for defence and such, maybe with some different possible aproaches that the network could pick...
And try to set up the individual bots so that mutations will affect the weights that they use.

I know there was some thoughts about allowing a bot to switch tie mid cycle, but not sure when or if it will be done. Buit this would definately help a lot. This is also why I'm sugesting a fungus like creature to begin with, since slow changes can work well for it. But you'll have to be tricky, I would sugest using .tieloc and .tieval to force values into the neurons, or something like that... but you would still need atleast some sort of syncronization to get it all working... I can't see an easy way of doing this.

Offline jknilinux

  • Bot Destroyer
  • ***
  • Posts: 468
    • View Profile
Neural Net MB?
« Reply #2 on: October 25, 2008, 07:18:41 PM »
Oh, I didn't know I could only manipulate one tie per cycle... That's awful. Why did we do that? Or is it just too complex and the suggestion to make it is still open?

Offline bacillus

  • Bot Overlord
  • ****
  • Posts: 907
    • View Profile
Neural Net MB?
« Reply #3 on: October 25, 2008, 07:58:35 PM »
I believe the discussion is still open.
"They laughed at Columbus, they laughed at Fulton, they laughed at the Wright brothers. But they also laughed at Bozo the Clown."
- Carl Sagan