well let's see what the behavior it's
serve the best individual

at different generations. At generation 1

that is the very first generation where
we used random strategies

the population was just a bunch of
random strategies the best

average score was -81. Let's see why

that's true here's the behavior that
strategy

on this particular environment of cans

it's the individual moves down and

then it crashes into a wall then it crashed
into all again

and again, and you can guess the rest of
its life

its gonna spend crashing into a wall
scoring

-5 points each time crashes

So this slightly different scoring
different environments but

on average it does very poorly. Okay by
generation 10

the best average score was all the way
up 0

okay that's what better than -81. What
did that individual do

well here's that strategy on this
particular configuration of can.

And here's what it did.

In other the words it did nothing.

he just stayed put. By staying put

it didn't crash into any walls or do any
other bad things

and so it got a score of 0. Well that's
not very a

satisfying but this was better than the
other individuals in its population

therefore it got to spread streets more
than

other individuals its generation. By
generation 200

the fitness was up to 240 for the best
individual.

Here's what that did. It moved down

doesn't pick up the can but moved down
further that picks up the can

then it comes back up pickup a can.
Now it's gonna move back down

moves all the way down to the bottom the
cans and then picks them up

so this one's doing a lot better than
the previous one

that long we saw it generation zero this
course 240

generations later but this individual

is picking up cans but tournament and
inefficient way

she could see and by generation a
thousand which was

the a maximum I let it run fitness is
492

on average and here's what this
individual did.

It does exactly what you'd expected to
do it moves down,

systematically picking up cans. My
strategy

remember said if you're have a can
your site

pick it up otherwise if there's it can
adjacent site move there and so on

this one is systematically going down
and up

and down and up. it's more systematic
than my strategy

Now there is one improvement on my strategy
But there was actually another

improvement which is a very

surprising to me: innovation on the part
of the genetic algorithm

to explain why the genetic algorithm

outperformed mine, I looked at

my strategies behavior compared with the
genetic algorithm strategies behavior in

many different

environments. Here's one particular type
of environment that stood out.

So here's Robbie in a cluster of cans

so my strategy says pick up a can
in the site if you can

then move to an adjacent site if there's
a can in an adjacent site

we're here this cans into adjacent sites
so picks one at random

decides to move west. okay it picks up
the can

but now because it cannot see

past one square in either direction

it can't see this can overhear since it
has no memory

it's lost all information about this can

and it stuck it doesn't see any cans
around it and doesn't remember that

had cans over here. However the genetic
algorithm

figured out a way to get around that
problem

and here's how it did. This is a genetic
algorithm strategy

it says don't pick up the can

it moves over to the West

now it picks up that can but now because
it did not pick up

this can here, it has sort of breadcrumb
trail

or tray love cans to get back

to wear the cans cluster cans were

so it doesn't have any memory itself
probably doesn't have any memory

but botanic algorithm invented this kind
of external memory by leaving cans there

and then having the robot come back and
pick them up

so it was pretty ingenious I thought

and that's an example other genetic
algorithm actually thinking of something

that

I didn't think of thus doing better

then my own strategy did.

Robbie is a robot is a very simple
example that

I came up with just to teach people
about genetic algorithms

but it does illustrate some interesting
and general principles of evolution

that are often seen in genetic
algorithms. Let me just say what these

are

First is that natural selection works

we went from a bunch of random
strategies that perform very poorly

and just using this kind of natural
selection

inspired by Darwinian natural selection

the genetic algorithm evolved what is
essentially a perfect strategy

also evolution in our system

seem to proceed be a period serve stay
sis

that is stain at about the same fitness
level

punctuated by periods of rapid
innovation

and that something is also seen in
biological evolution

in both molecular evolution for since
evolution of viruses

and evolution a bacteria all through to

large-scale evolution thats seen in the
fossil record closes a little bit

controversy on

in general were able to show that in
genetic algorithms

this kind of behavior is quite common
even though there's no

external events happening to cause this

rapid steep innovation.

And I should say that this is in
biological evolution

this kind of phenomenon has been called
punctuated equilibria.

another principle is that the phenomenon
of accept haitian is common

acceptation means the shifts in a
function of a trait

during evolution and we saw that
happening

in that trait that cost the robot

to not pick up a can but move past it

So you actually saw that happening at
generation 240

and that was the inefficient version of
the strategy.

there it wasn't a very good treat
because it cost the robot to be very

inefficient not pick up all the not have
time to pick up all the cans

but it was hold by the genetic
algorithm by evolution

to actually be a very efficient trait in
that it became only used in

appropriate circumstances like the one
I just showed you

where I compared my strategy to the
genetic algorithm strategy

and so that trait was shaped by
evolution

from being a rather non adaptive

to a very adaptive trait and that's also
seen in biological evolution.

And finally the dynamics and results of
evolution are somewhat unpredictable and

hard to analyze

even in this extremely simple version of
evolution

and it takes a lot of work to try and
understand what

though results a pollution are doing why
the

are fit had how they evolved these are
all

in genetic algorithms unpredictable and

typically either very hard or impossible
to analyze.