Sorry, an error occurred while loading the content.

## lilgp (or other gp) in real-time?

Expand Messages
• Hi, I have made extensive modifications to lilgp for a project I am playing with, but am having problems with this latest bit. I sent an email to someone and
Message 1 of 2 , Oct 15, 2002
Hi,

I have made extensive modifications to lilgp for a project
I am playing with, but am having problems with this latest
bit. I sent an email to someone and would like that I had
kept a copy of the email, but oh well.

Take for example an environment, two dimensional for
simplicity, that has scattered around a biotoxin. Givenn
the cartisian coordinates of the biotoxin generate an
individual that evaluates to a probability that a given
coordinate has biotoxin at that location. This problem
works for static locations of biotoxin. Now say that this
two dimensional space is a river, still only dealing
with the surface for simplicity, and that river moves
a given distance X in a given time unit T. Take the gp,
'prime' the individuals with the static data so the gp
has a place to start, and then start the river moving.

For the moving window of data, add 1X of data to the
beginning (most recent) side of the data window and
remove 1X from the end (least recent) side of the
data window. Like the window of a train watching the
landscape roll past.

So far I have reduced the overall distance of the river
analyzed from 100X to 10X so the gp is more responsive
to new data entering. But, the gp is not responsive
enough. Using lilgp, which I like a lot, I have modified
the run_gp() function to recalculate the adjusted
fitness of each of the best individuals each time new
data is added to the to the data window, but it seems
as if the overall idea is not working. It seems as if
the gp (lilgp) is still keeping the absolute (best of
run) adjusted fitness regardless of where I re-evaluate
the equations for the data.

Can anyone help?

I am close to writing my own gp to make this work,
but for the time involved and that lilgp has so many
things about it that I like... I'd rather not write
my own.

Mike
• Hi Mike, lilgp caches the fitness values of the genomes for efficiency. If the fitness will change over the lifetime of the individual, you need to invalidate
Message 2 of 2 , Oct 15, 2002
Hi Mike,

lilgp caches the fitness values of the genomes for efficiency. If the
fitness will change over the lifetime of the individual, you need to
invalidate the cache each time. In the app_eval_fitness routine, you need
to do this:

ind->evald = EVAL_CACHE_INVALID;

Hope this helps,
Terry

On Tue, 15 Oct 2002, Mike Egglston wrote:

> Hi,
>
> I have made extensive modifications to lilgp for a project
> I am playing with, but am having problems with this latest
> bit. I sent an email to someone and would like that I had
> kept a copy of the email, but oh well.
>
> Take for example an environment, two dimensional for
> simplicity, that has scattered around a biotoxin. Givenn
> the cartisian coordinates of the biotoxin generate an
> individual that evaluates to a probability that a given
> coordinate has biotoxin at that location. This problem
> works for static locations of biotoxin. Now say that this
> two dimensional space is a river, still only dealing
> with the surface for simplicity, and that river moves
> a given distance X in a given time unit T. Take the gp,
> 'prime' the individuals with the static data so the gp
> has a place to start, and then start the river moving.
>
> For the moving window of data, add 1X of data to the
> beginning (most recent) side of the data window and
> remove 1X from the end (least recent) side of the
> data window. Like the window of a train watching the
> landscape roll past.
>
> So far I have reduced the overall distance of the river
> analyzed from 100X to 10X so the gp is more responsive
> to new data entering. But, the gp is not responsive
> enough. Using lilgp, which I like a lot, I have modified
> the run_gp() function to recalculate the adjusted
> fitness of each of the best individuals each time new
> data is added to the to the data window, but it seems
> as if the overall idea is not working. It seems as if
> the gp (lilgp) is still keeping the absolute (best of
> run) adjusted fitness regardless of where I re-evaluate
> the equations for the data.
>
> Can anyone help?
>
> I am close to writing my own gp to make this work,
> but for the time involved and that lilgp has so many
> things about it that I like... I'd rather not write
> my own.
>
> Mike
>
>
> To unsubscribe from this group, send an email to:
> genetic_programming-unsubscribe@yahoogroups.com
>
>
>
> Your use of Yahoo! Groups is subject to http://docs.yahoo.com/info/terms/
>
>
>
Your message has been successfully submitted and would be delivered to recipients shortly.