Searching \ for '[PIC]: Genetric Programming' in subject line. ()
Make payments with PayPal - it's fast, free and secure! Help us get a faster server
FAQ page: techref.massmind.org/techref/microchip/devprogs.htm?key=programming
Search entire site for: 'Genetric Programming'.

Exact match. Not showing close matches.
PICList Thread
'[PIC]: Genetric Programming'
2001\10\17@194318 by James Caska

picon face
>I have had good results with GAs and populations of 100 with only a
>moderate number of generations (<50).

I agree on two accounts

1. You can get good results with small populations if the problem is small
enough.
2. Yes it is difficult to write simulations... but it is a matter of
constaints

The question is? Which are you going to build. A difficult control problem
with simple mechanics or a simple control problem with complex mechanics. Is
it very interesting to solve a simple control problem using a GE/P?/
Perhaps. On the other hand, with a bit of thought, a simple simulation might
be discovered to represent a complex control problem. The complex control
can be discovered virtually, the mechanics built only once.

For the physical robot, unless it was highly constrained such that the
control solution was all but trivial it would be very difficult to predict
success. Nevertheless I am interested in being involved.

James

{Original Message removed}

2001\10\18@062800 by mail

flavicon
face
James Caska wrote:

> -----Original Message-----
> >I have had good results with GAs and populations of 100 with only a
> >moderate number of generations (<50).
>
> I agree on two accounts
>
> 1. You can get good results with small populations if the
> problem is small enough.

Every problem is unique. I think it's not so much about the size,
necessarily.
It may take some tinkering to find out what population size is ideal for
the problem at hand.


> 2. Yes it is difficult to write
> simulations... but it is a matter of constaints

Do you mean, constraints in the problem, or constraints as in
time/money?

Eventually, something needs to be built and might need the ability to
learn.
Problems can be simulated in advance (and should be).
It would be useful imo to have an architecture that enables evolution
(or co-evolution).
(And indeed, it would be most useful if it was an open architecture!)

FPGAs may be better suited to the task than PICs. I'm just getting
started in embedded dev, and the PIC is my choice. Can't wait to do
something with GA or GP outside of a PC :-)


>
> The question is? Which are you going to build. A difficult
> control problem with simple mechanics or a simple control
> problem with complex mechanics.

Firstly, my motto is KISS.
Secondly, I would rather solve problems in software than in
hardware/mechanics (easier to change, plus I'm a software guy...)

Genetic Algorithms have their use, but if a simpler solution exists the
choice is obvious.

However, having an 'Evolutionary platform' or whatever it should be
called will definitely have its use.


> Is it very interesting to
> solve a simple control problem using a GE/P?/ Perhaps. On the
> other hand, with a bit of thought, a simple simulation might
> be discovered to represent a complex control problem. The
> complex control can be discovered virtually, the mechanics
> built only once.

Many control problems can be solved via simulation.
But what if the environment changes in unexpected ways?

At some point, something needs to be built. (Can't paint real walls in a
simulation!)
Ideally, it can improve itself from that point on and adapt to
unexpected changes in environment.

Having it solve problems we already know how to solve is a good starting
point, I think (verification).

> For the physical robot, unless it was highly constrained such
> that the control solution was all but trivial it would be
> very difficult to predict success.

True. But in GA/GP we don't really predict it, we measure it (although
we'd like to predict it).


> Nevertheless I am
> interested in being involved.

Good!
I don't know when I will be up to speed enough with assembly (and
electronics...) to get started with GA on a PIC.
My initial question was mostly a query to see whether in principle it
would be possible.
I will start thinking about a general architecture and some practical
applications in the mean time.

Peter.

--
http://www.piclist.com hint: To leave the PICList
spam_OUTpiclist-unsubscribe-requestTakeThisOuTspammitvma.mit.edu


More... (looser matching)
- Last day of these posts
- In 2001 , 2002 only
- Today
- New search...