HOLY! I swear I understood some of that! *brain explodes now (but head is okay)*

Btw, if you have trouble with the limit, add some text files along with it. That way, you won't have to double post, and you could probably even shorten posts.

OK!!! I had a lot of stuff to do in the meanwhile but what I did do was implement a couple of experiment in Breve.
Guess which physics engine breve uses? ODE, the open dynamics engine, which is the same as Toribash. I'm going to implement my own rig of Tori later and test the fitness functions based on a few things. Later on I'll be trying to get it working with EANT and hyperNEAT/CPPN but here is a taste of things to come!!!!!!

(NOTE: this is just tests using some custom configurations of blocks and joints; specifically I made a pogo stick jumper guy and a "snake" or "worm" by connecting 5 blocks with 4 joints. eventually I will be attempting to get a 1-1 correspondance with toribash so that the AI will directly be able to plug in. Keep checking back for updates.)
worms:
http://ca.youtube.com/watch?v=WnjsermaVi0
http://ca.youtube.com/watch?v=S1GZGXd7iyo
http://ca.youtube.com/watch?v=hhTSJJDew4Y

hoppers:
http://ca.youtube.com/watch?v=-3TRIFF3xJE
http://ca.youtube.com/watch?v=PgHqQ-Szy70
http://ca.youtube.com/watch?v=C_fwrXNTzVM

These are all examples of evolution happening on a single computer over a couple of hours; when I have the full thing completed how I want it, it should be able to be networked across many computers and produce far more general and human-competitive results.

Hi Tripwire,

I was just doing a search for implementations of EANT2 on the internet and I saw your post.

Would I be able to get a copy of the source code? I am doing quite a bit of work with NEAT and I would like to compare the two.

I don't want to promise anything yet. But if I do a lot of work with the code. I could give you any modifications I make to the code. Or we could even GPL it.

Thanks for your time,

Shane

Hi Shane, its good that you are interested in EANT2, however right now I haven't even come close to implementing it. Since I work and study I only get time to program on hobby stuff infrequently; right now I'm doing my work with python-neat, written by Cesar Gomez. He has GPL'd his source code and it is available at http://code.google.com/p/neat-python/
If you want to help me implement EANT2 I'd gladly welcome your help.
While a work in progress the code for python-neat is very usable and contains a lot of interesting features. HOWEVER, it is using NEAT which is quite outdated now; since its GPL I was planning on working in CPPNs/EANT/hyperNEAT in when I have time but to be honest I haven't done much but sketch out ideas yet. Additionally, the interface that I am using to Breve (the simulator which incorporates the open dynamics engine) isn't optimized at all; python simulations tend to be orders of magnitude slower than simulations coded in Breve's native scripting language; additionally Breve is single threaded and doesn't support exporting objects as xml yet which makes it a bit more annoying to save your work and pick up where you left off. Luckily these features have been incorporated into python-neat by Cesar so from that end you are safe.

The thing that really prevents python-neat being able to work for this application right now is that NEAT is too dumb an algorithm; starting out with fully connected nets (which I think is pretty stupid in this case) we have over 820 connections at start; this means that in the first generation you are trying to optimize an 820 variable equation simultaneously. Not the best way to go if you want to finish evolving something anytime this millennium. The way forward is obviously incorporating all the best new breakthroughs like EANT2's minimally structured nets, and superior CMA-ES weight optimization. Exploiting task geometry with hyperNEAT also sounds promising but I can't visualize exactly how it would be implemented into a physical controller; if you have any ideas please tell me.

I've just finished making my first toribash-like simulation experiment for evolving a walking controller. I will work on making this more close to how Toribash operates and then release the experiment files for people to use if they wish.
Heres a screen of it trying to balance

Great Work for now TripWire ...
Im watchin this project closely , if finished this will be something unique
in the toribash history lol ;]

Hi Tripwire,

It might be interesting to chat. I am doing quite a bit of work with NEAT. Me and my colleague have implemented our own version of NEAT in C++. But we are having problems choosing optimal NEAT parameters. Which is why EANT2 is so interesting.

I happened to notice you posted to the Canadian YouTube. Are you Canadian? I am from Victoria originally but I live in Dublin Ireland now.

In any event I'd be interested in chatting with you a bit. If you get a chance drop me a line at: shane dot mcdonald dot ryan at gmail dot com

In the coming months implementing our own version of EANT2 is a distinct possibility. I'd welcome the chance to pick you brain on the subject.

Regards,

Shane

Some early best-of-run nets: http://ca.youtube.com/watch?v=myqIEMO4R3I
They showed up around generation 24 with population size of 400 (i.e. it took more than 9600 evaluations to get to this level of taking two steps and then having a seizure. God knows how many NEAT would require to get natural looking movement, if it is capable of doing so.)

The fact that they were the best in their populations shows you how much poorer the rest of the population was (most of them fell over, spastically twitching as soon as the simulation starts). This is an indication that a) NEAT as expected is very slow, especially on huge nets, and b) my current simulation settings aren't conducive to making natural motions, at least not when you start with each net output being connected to 42 inputs. I'd really love to get some feedback from Hampa on how he implemented Toribash (specifically the joint parameters used). If I set any joint damping at all I tend to make the simulation incredible unstable; same goes for setting joint torques directly (right now I'm using the joint-motor function to set target velocities for joints, with strength limits taken from the default.tbm file.)

how about a script where uke senses the nearest body part and trys to block you

call it blocking.lua

Heres a video if it doing a sideways kind of flip
http://ca.youtube.com/watch?v=un5lSsiqQNs
(Keep in mind that this was purely accidentally: the fitness criteria was the amount of time before it fell over, and in one of the early generations the best solution just so happened to be this awkward flip. Later generations have all sorts of hopping around on one foot and tumbling over maneuvers. )

That's some serious drunken kung fu. The flip is actually a pretty nice kick.