Subject:
|
Re: Design
|
Newsgroups:
|
lugnet.robotics
|
Date:
|
Wed, 7 Dec 2005 19:10:43 GMT
|
Original-From:
|
dan miller <danbmil99@yahoo!nomorespam!.com>
|
Viewed:
|
1710 times
|
| |
|
|
--- steve <sjbaker1@airmail.net> wrote:
...
>
> A good example here (in an admittedly 'toy' world) was of a guy who
> wrote a careful simulation of a simple 'world' and had creatures made
> of blocks that were formed from 'genes' and brains that were made from
> simple subroutines - also formed from genes.
>
> He set up his world - setting up his 'evolutionary' system to allow
> the creatures that moved the fastest to survive and to breed the next
> generation - whilst the slowest creatures were killed off. His
> criteria was to race the creatures. The first to cover a certain
> distance (or the one that get closest to the line if neither makes
> it within a certain time) was the winner.
>
> His first effort ran for a bazillion generations.
>
> He'd hoped for some kind of interesting locomotion to 'just evolve'.
>
> What actually happened was that his creatures evolved into enormously
> tall blocks with no brains whatever. All they did was to fall over.
> However, being so enormously tall, when they fell over, they crossed
> the finish line - and thereby out-evolved any possible elegant walking
> machine!
>
> He patched up his algorithm in a variety of ways - but each time,
> his creatures would evolve into some totally un-interesting form
> that would win by exploiting loopholes in his testing scheme.
>
> He even ended up with creatures that collected energy by exploiting
> bugs and roundoff errors in his physics simulation! For those
> virtual creatures, roundoff error was as natural a resource as
> grass is to cows...so they learned to exploit it.
This is fascinating stuff -- did the fellow publish? Any pointers to his
work would be greatly appreciated.
It ties in directly with some research I've been doing on reversible physics
and computation, and some ideas I have about evolutionary programming,
a-life, etc. My basic thesis is this: you can't scale up solutions crafted
for toy worlds; what you need to do is scale up the *process*, so that the
toy world is equivalent to, in all the aspects that are important to you,
the real world. Reversibility is one example: what it really means is,
conservation of information; in a physics simulation, that translates to
proper conservation of energy. This allows your creatures to have an
'energy economy', which should in theory mitigate the problems your friend
was having.
There are many other details, but the general idea is this: create a
simulation that is realistic enough, and it will no longer be a 'toy' world
-- it will be a realistic model of the real world. Solutions crafted for
this virtual world (whether by hand, evolutionary methods, or a combination)
will then have a good chance of being transferred to our world successfully
(there's more to say about this, such as proper introduction of noise and
error -- but later if anyone is interested).
The best analogy I've been able to come up with is the Wright brother's use
of wind tunnels. Prior to that development, research into heavier-than-air
flight was pretty much hit and miss; no one had gotten a firm enough grasp
of the fundamental parameters of the design problem. Once they got through
that critical juncture -- in part by developing a technique that increased
their research productivity by an order of magnitude -- progress was made
quickly. And the rest is history.
-dbm
__________________________________________
Yahoo! DSL – Something to write home about.
Just $16.99/mo. or less.
dsl.yahoo.com
|
|
Message is in Reply To:
 | | Re: Design
|
| (...) There are many answers to that. One is that evolution *IS* used in some fields of engineering. Genetic algorithms have been used in many situations. I myself have used a genetic algorithm to develop a program that can recognise buildings in (...) (19 years ago, 7-Dec-05, to lugnet.robotics)
|
5 Messages in This Thread:
  
- Entire Thread on One Page:
- Nested:
All | Brief | Compact | Dots
Linear:
All | Brief | Compact
|
|
|
|
