W. B. Langdon and R. Poli
----------------------------------------------------------------------
Date: Mon, 30 Jun 1997 09:49:21 +0100
To: Bill Langdon <W.B.Langdon@cs.bham.ac.uk>, R.Poli@cs.bham.ac.uk
From: Chris Gathercole <chrisg@harlequin.co.uk>
At 18:38 28/06/97 +0100, you wrote:
...
> You are right the question of whether to suppress bloat or not is a
>very interesting one. There is an argument that some form of
>redundancy is in practise necessary for long term evolution.
Yo Bill and Riccardo,
I liked your "GP with One-Point Crossover" paper. I'm envious. It sounds as
though a simple idea suddenly exploded into lots of interesting aspects. My
impression is that the problem (or is it actually a problem?) of tree-bloat
is under attack from all sides.
On the theme of bloat, and in response to your msg on 'fitness causes
bloat', I thought it worth putting in my ++2p. I've dabbled with adaptive
fitness functions (i.e. ones which adjust on the fly to the current
performance of the population or best-of-gen), and two of the more
interesting things that happened are:
- that it allows the runs to go on effectively for many many more
generations, - and even with weak parsimony (i.e. the parsimony
contribution to the fitness function is smaller than the smallest unit of
fitness, i.e. it only distingishes between two trees which would otherwise
have an identical fitness), the best-of-gen and avg tree sizes are much
larger than with static fitness functions and parsimony. Plucking a figure
out of the air, I would guess that the adaptive fitness fns I've tried
result in trees which are 3x larger, only ever shrinking after an optimal
tree had been found. The population never has time to prune the best-of-gen
trees because they don't last long enough as best-of-gen trees. The
shifting fitness function penalises the current best-of-gen trees by
reducing the importance of the things they do well.
I can't yet say how the AFFs would perform without parsimony, i.e. with
unconstrained tree sizes, I have a suspicion that they might bloat, but I
could be wrong.
I'll be trying to say something along these lines at GP97.
-Chris