The details of the algorithms used by Karl Sims

to evolve the virtual creatures are quite complicated

and I'm not going to go over all the details. I'm just

going to give a very sketchy outline. This part

is optional - it's only if you're really interested

in hearing about how this is done. You can also

read Karl Sims' technical papers on this topic by

going to the link that I've put on the course

materials page. As I mentioned, the algorithm

evolves both the morphologies of the bodies and

the neural networks that control the bodies.

The creatures' bodiy morphologies were represented

as these grammars which are repeatable little programs

that can produce repeated structures like these. The

brains, or the neural networks, or the networks of

neurons - the neural network had an input of

sensor values and it output effector values which

were forces on the joints. The sensors included:

joint-angle sensors - that is, each joint reported

the value of its angle, - contact sensors - which

were 1.0 if a contact was made with anything and

-1.0 if no contact, and photo sensors which returned

the coordinates of light sources relative to the

orientation of the part. So as you can see, it's getting

kind of complicated, and of course, these were all

virtual. These are not real robots - they are all computer-

simulated. The creatures' brains...each neuron

computes its output value from its inputs. The genetic

algorithm is going to evolve the neural network

and the output function of each neuron. The

system had effectors, which you can think of as

muscles where each effector controlled a degree

of freedom of a joint and each effector gets its

input from a single neuron - or a sensor - and outputs

the force to put on the joint. So here's a picture

from one of Karl Sims' papers showing the whole

system with the genotype which is evolved

to produce a phenotype - the body. The neural

network is also evolved to produce a phenotype,

so these are simultaneously evolved - the body

and the brain. Inside of this physical simulation which

is implemented in parallel on what's called a

connection machine which is a massively parallel

machine. And as you can imagine, this whole

algorithm takes a lot of computation. Here's how

the creatures are evolved. The creature is grown from

its genetic description and to calculate its fitness,

it's run in simulation for some period of time.

Okay, so the sensors provide data about the world,

the brain produces effector forces to move the

creature and the fitness is "how successful was the

desired behavior such as swimming or hopping

or following a light or competing" by the end

of the allotted time. So that's all for my quick overview

of this extremely complicated algorithm, but

I'm sure you agree that the results are a lot of fun to

watch and very thought-provoking.