Re: Particle Swarm Optimization

In article <88363612213559@[EMAIL PROTECTED]
>, Marcel Hendrix <mhx@[EMAIL PROTECTED]
>
wrote:
>Albert van der Horst <albert@[EMAIL PROTECTED]
> writes Re: Particle
Swarm Optimization
>[..]
>
>> Was your ampliphier a discrete problem?
>
>No. But it has multiple solutions at (sub)multiples of the
>switching frequency, and near a solution the phase relation
>between current and voltage is extremely im****tant (ns resolution
>required). Furthermore, I needed to minimize a vector function,
>so it was necessary to 'play' with the weights of the
>vector-to-scalar conversion.
>
>[..]
>
>> Given there are about thousand products a week, so the search space
>> was 1000! to start with, and given that computer assisted planning
>> already was very near optimal, I gave the customer little hope that a
>> s.a. solution could better this, so the project was cancelled.
>
>How did you know it was already optimal? (I assume you not really
>meant to say that was a given). Even if your client would not
>implement the final program, they should've been interested to
>know by how much their current software was off the optimum.

Remember this is a steel plant, where you want to use up the casting
time, and use up walzes. You don't want to produce waste, and you
try to produce only what is ordered. On all these points production
plans scored pretty good.
In the second place optimal was not exactly defined. So there were
some pretty large error bars around evaluation functions. This
could have led to solutions that were considered good but actually
weren't. Not that we got nearly that far.

>
>> But this left me wondering how far s.a. could have got with this
>> problem. Evaluating a planning involves running a simulated
>> factory over it, which is much more costly than a step evaluation
>> in a typical travelling salesman problem.
>
>For the amplifier problem, a single algorithmic step costs a full
>eight HF-cycle simulation of the entire circuit. I didn't do the
>simulation in SPICE, but wrote the circuit equations in Matlab
>so that the amplifier became a function of its component values.
>If I need to take it further (the project is by far not finished
>yet), I'll do only the inner steps with SPICE (sort of a
>space-mapping approach).

Can you give an idea of the search space? How many variable
had to be tweaked, and to what precision?
Did it solve the problem in split-seconds, seconds, or hours?
I read that simulated annealing takes on the order of thousands
evaluation points at least.
>
>-marcel
>


--
-- 
Albert van der Horst, UTRECHT,THE NETHERLANDS
Economic growth -- like all pyramid schemes -- ultimately falters.
albert@[EMAIL PROTECTED]
 &=n http://home.hccnet.nl/a.w.m.van.der.horst