DPRG
DPRG List  



[DPRG] Great Roborama!

Subject: [DPRG] Great Roborama!
From: Francisco Jose Ayala ze at neuroblast.net
Date: Wed Nov 26 09:23:55 CST 2008


> Hi Jose,
>
> Can you share more on how you develop the neural nets, do the  
> simulations, and genetic algorithms?

Be glad to.  I discussed what the genomes look like and a bit about  
how they evolve in my previous email response to Jacob.  Here I'll add  
a few more points.

The neural systems (I call them brains) that we develop have very  
little in common with conventional artificial neural networks (ANNs).   
ANNs consist of layers of neurons that each perform identical  
threshold gating functions, and the layers are fully interconnected,  
meaning that every neuron in a given layer is connected to every other  
neuron in the layer ahead of it and behind it.

By contrast, the neurons in our brains are free to evolve almost any  
kind of mathematical function, each neuron in a given genome is  
usually quite different from every other, and the connections between  
them are free to evolve just about any kind of pattern.  You can see  
what one of our earliest and simplest brains (which we evolved to  
balance an inverted pendulum) looks like, along with a detailed  
description of how it works, here.

A genome consists of genes (encoded as parse trees) for multiple  
neurons.  A brain consists of multiple instantiations of that set of  
neurons.  A population consists of multiple brains.  Every generation,  
every brain within a population is evaluated on its performance on a  
given task (e.g. line-following).  The brains that perform best stand  
the best chance of contributing their genes to the gene pool of the  
next generation.


> Is this software proprietary to you or available from others?

Our software and evolutionary system is proprietary.  We normally do  
not make it available to others, but we would be willing to discuss  
sharing with fellow DPRGers.  If you have any ideas for cool robotics  
(or other) applications that could use this kind of AI, let's talk  
about it.


> I have had an interest in Genetic Algorithms and Programming for  
> awhile.

My background is in molecular evolutionary genetics, so naturally this  
heavily influenced my approach to AI.  The only design paradigm to  
ever produce a genuinely intelligent system is biological evolution.   
And the only kind of intelligent system it every produced was neural  
(that is, simple interactions among simple processing elements, from  
which potentially emerges great computational complexity).


> Also have done some work with the NASA CLIPS expert system.

Very cool.


> I gather you are not training the neural network per se, but using  
> GA to derive networks and then test in simulation?

Right.  There are no training sets, backpropagation, descent  
algorithms, etc.  The brains either perform well or the selection  
regime will purge them from the population.  To paraphrase Yoda: Do or  
do not; there is no train.

That is not to say, however, that we cannot evolve brains that learn.   
In fact, we are working on evolving brains that learn through operant  
conditioning, so their behaviors can be taught, much like you might  
teach tricks to a puppy or a dolphin.

This opens up the possibility for some interesting robotics  
applications.  Again, if any DPRGers have any ideas, let's talk.

Cheers,
Jose'



> Rud Merriam K5RUD
> ARES AEC Montgomery County, TX
> http://TheHamNetwork.net
>
> -----Original Message-----
> From: dprglist-bounces at dprg.org [mailto:dprglist-bounces at dprg.org]  
> On Behalf Of Francisco Jose Ayala
> Sent: Monday, November 24, 2008 4:28 PM
> To: dprglist at dprg.org
> Subject: Re: [DPRG] Great Roborama!
>
> Yes, our robots, 3piRoBrainia I and II, were driven by AI.
>
> Specifically, they were controlled by a neural system that we  
> evolved in a genetic algorithm-like virtual world of genes, genomes,  
> organisms and populations that all compete against one another in a  
> rigorous selection regime.  The end product is a "brain" composed of  
> many identical subunits (just as a human cortex is a massive array  
> consisting of countless repetitions of the same assemblage of  
> neurons).
>
> The 3piRoBrainia brains consist of five such subunits, one for each  
> reflectance sensor.  You can see a picture of this brain here:  http://www.neuroblast.net/images/20081121pop17gen690hiBrn5.jpg
>
> Each collection of blue circles is a parse tree that defines the  
> processing of a single neuron.  The lines represent neural  
> connections, or axons: pink lines are for connections within a  
> subunit, and green lines for connections between adjacent subunits.   
> Yellow circles represent the location of sensor inputs.
>
> We often have no idea how these brains actually work; we simply  
> select for certain capabilities, such as line following, and then  
> load the resulting brains onto the robot.  The brains get better  
> with every generation.  The brains we entered into the roborama  
> competition were produced in just 48 hours of evolution.  We'll  
> bring improved brains to an upcoming RBNO.
>
> Also, Dale Wheat is creating a modification of the 3pi for us that  
> will have a 2D array of 3x8 sensors, which will give the brains a  
> much better picture of the road ahead.  We'll have brains for that  
> soon.
>
> Cheers,
> Jose'
>
>
> On Nov 24, 2008, at 8:34 AM, DeltaGraph at aol.com wrote:
>
>> Dale,
>> Maybe you can elaborate?
>>
>> Was that Jose Ayala and Carl Sturner's 3pi robots?
>> I heard brief mention of a "Brain".
>>
>> Their robots were right up there in the fast run times.
>>
>> Ron
>>
>

-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://list.dprg.org/pipermail/dprglist/attachments/20081126/c04e5457/attachment-0001.html

More information about the DPRG mailing list