Messages

1

Off Topic / Open Source evolutionary simulation of social bots

« on: October 04, 2010, 04:47:51 PM »

Hello everyone, this is a follow-up to a thread I posted here a while ago:
http://www.darwinbots.com/Forum/index.php/topic,3382.msg1382424.html

It was about a simulation that is similar to DarwinBots, but has some different I/O conventions and brain dynamics.

I finished porting the simulation (originally done in Python) to C++. It now runs MUCH faster, and can handle many more agents. I also made it Open Source and posted it on google code
http://code.google.com/p/scriptbots/

Some predator prey plots I usually get:


I also made a short video just talking about it for a bit:
http://www.youtube.com/watch?v=qSmlKAly1UE

I'm posting here because I thought some of you could be interested. I was pretty happy with some of the behaviors I was able to observe, on a high level.

There are two discrete types of agents now: Herbivors and Carnivors. Herbivors eat food, and Carnivors eat Herbivors. I was able to observe individual hunting techniques and good steering in carnivors. They also learn pretty quickly to avoid things of the same color (which are to high degree of confidence their descendents). For herbivors, I saw many that learned how to escape from Carnivors, come up with manuvers that seemed to be designed to fool carnivors (go to left and then sharply turn right and boost), and behavior where herbivors started to impersonate carnivors by emitting the same color that carnivors did.

I was unable to observe too much coordinated behavior, such as hunting in packs, or something like that. I believe this to be way out of the scope of the project because of the low amount of sensors, the sizes of the brain, and other environmental variables such as length of a life, etc.

I think a much better environment to compare to would be some bacterial life forms. Does anyone know of animals, or systems that have predator prey dynamics, but are very simple? Maybe some worms, or fishes?

Any suggestions of other things I could be looking for, or other simulation mechanics I could develop to get them to be more social? other types of brains?
Cheers!
-meph

2

Off Topic / My simulation: Evolving social bots

« on: May 03, 2010, 08:48:36 PM »

I agree with what has been said.

The use of Neural Network is loosely inspired by the brain. (But way too loosely )

Its interesting to experiment with brain modules. Interestingly I also tried Radial Basis function networks and they perform MUCH worse.
I also tried implementing more higher level neuron types... so some neurons would not compute sigmoid, but do something like a MAX() of inputs, or StandardDeviation() of inputs, or they would be a little if_then_else module working with inputs. Those brains are therefore more quickly higher in the hierarchy of capabilities. Sigmoids are universal approximators so you could do all of this with simple perceptrons but may just need too many of them.
It is interesting by the way that the above idea did not really work too well. I think maybe these brains are harder to "train", I'm not sure. Just plain perceptron brains worked best so far. (weird)


Anyway, my current goals are to change the simulation to something more along the lines of DarwinBots, where the brain is a "program" and see what i get. I will also make the code nicer and maybe release it sometime in the future. My schoolwork has just finished, so I can start working on this now

Thank you for comments!
cheers
meph

3

Off Topic / My simulation: Evolving social bots

« on: April 29, 2010, 01:10:24 PM »

Looks pretty cool   Have you gotten anything like a stable ecosystem going?  (eg: predators and prey following a sinusoidal graph).  That's always been the hardest part in Darwinbots (it's been done, it just requires a lot of tweaking to get the initial parameters set up right).

Hi: well that's just the thing, I can't make those plots because I don't have this distinction. These bots just do their thing (whatever that is), and you can only qualitatively decide, sometime. Many times it is clear, but many times they are in this grey area in between where they are gatherers but then flip randomly and attack something.

If Darwinbots does have this clear distinction where every bot must be one or the other, then I would consider that as an instance of where you are putting in too much structure. You are baking the solution in, not seeing it emerge... which takes some fun out of it for me

4

Off Topic / My simulation: Evolving social bots

« on: April 25, 2010, 05:42:19 PM »

Hello, I just thought some people here would be interested in seeing a simulation that I put together a while ago on this subject:
http://www.youtube.com/watch?v=GvEywP8t12I



I'm not too familiar with darwin bots but there are some differences that I picked up on right away.
I guess the biggest one is that I do not use a coding-based DNA. Instead, my DNA specifies the structure, connectivity, and parameters of the perceptron network. My brains are therefore automatically robust to small mutations because of the invariance in the sigmoidal activation function.
However, compared to Darwinbots I think I am working on a much lower level and can probably accomplish less elaborate tasks per parameter.

The sensory modalities are also a little different: my bots can "smell density of bots" and "hear speed of bots", and also sense the colors, but on a little lower resolution.

Oh and I don't have pretty graphs

I should check out darwin bots sometime
Cheers,
meph