source: branches/OKBJavaConnector/ECJClient/src/ec/parsimony/README @ 9674

ECJ supports several bloat control techniques.  Many of these techniques
are compared in detail in "A Comparison of bloat Control Methods for
Genetic Programming" by Sean Luke and Liviu Panait.  In this directory
we have implementations of several of them.

There are several methods in the article which aren't here.  The two you
should be aware of are BIASED MULTIOBJECTIVE, where we do multiobjective
optimization with fitness as one objective and size as another; and
plain-old LINEAR parsimony pressure, where the "fitness" F of an
individual is actually his real fitness R and his size S, combined in a
linear function, that is, F = A*R + S for some value of R.  We mention
these two because, like many of the techniques below, they perform well
over many different problem domains.  And importantly, LINEAR performed
the best in our tests!  Closely followed by RATIO TOURNAMENT SELECTION
(see below).  Double Tournament was pretty good too.  The problem with
linear parsimony pressure is that it could need to be tuned carefully --
though a setting of A = 32 seemed to work well in many problems.  Thus
you may wish to try out doing a simple linear parsimony pressure before
going to a more exotic method.  You can implement linear parsimony
pressure in your evaluation function: compute the fitness, then call the
size() method on the individual, then set the fitness of the individual
to the linear combination of the two.

Another technique commonly used to control bloat in GP is depth
limiting.  ECJ implements depth limiting Koza-style with a maximum depth
set to 17.  You can often get better results with a smaller depth limit. 
Interestingly, when the depth limit is set to 17, you can use depth
limiting in COMBINATION with ANY of the parsimony pressure techniques
discussed here, (including linear and biased multiobjective) and the
result is typically better than using them separately.



DOUBLE TOURNAMENT

Double tournament is a two-layer hierarchy of tournament selection
operators. Some N tournament selections are performed on some criterion;
and then the winners of those tournaments become contestants in a final
tournament.  The winner of the final tournament becomes the selected
individual.  You can have fitness as the first criterion and size as the
second criterion, or the other way around.

Here are good settings we've found for typical GP experiments.

[BASE] = ec.parsimony.DoubleTournamentSelection
# Do length as the initial tournaments, and fitness as the final tournament
[BASE].do-length-first = true
# The initial tournaments are of size 1.4
[BASE].size = 1.4
# The final tournament is of size 7
[BASE].size2 = 7

The default base is
  select.double-tournament


PROPORTIONAL TOURNAMENT

Proportional tournament is a single tournament selection; but some
percentage of the time the tournament is according to size rather than
according to fitness.

Here are good settings we've found for typical GP experiments.

[BASE] = ec.parsimony.ProportionalTournamentSelection
# The size of the tournament
[BASE].size = 7
# The probability that the tournament is by fitness (1.0 is equivalent
# to "regular" tournament selection).
[BASE].fitness-prob = 0.8

The default base is
  select.proportional-tournament



LEXICOGRAPHIC TOURNAMENT

Lexicographic tournament selection is simple.  We do a tournament
selection by fitness, breaking ties by choosing the smaller individual. 
Thus size is a secondary consideration: for example, if all your fitness
values are likely to be different, then size will never have an effect. 
Thus plain lexicographic tournament selection works best when there are
a limited number of possible fitness values.

Lexicographic tournament has no special settings -- it's basically the
same as plain tournament selection.  Here's how we'd set it up for GP
problems:

[BASE] = ec.parsimony.LexicographicTournament
# The size of the tournament
[BASE].size = 7

The default base is
  select.lexicographic-tournament



[DIRECT] BUCKETED TOURNAMENT SELECTION

Bucketed tournament selection is an improvement of sorts over plain
Lexicographic tournament selection.  The idea is to create an artificial
equivalency of fitness values, even when none exists in reality, for
purposes of lexicographic selection.  This allows size to become a
significant factor.  to create a set of N buckets.  The population, of
size S, is then sorted and divided into these buckets.  It's not divided
quite equally.  Instead, the bottom S/N individuals are placed in the
first bucket.  Then any individuals left in the population whose fitness
equals the fittest individual in that bucket are *also* put in that
bucket.  Then the bottom S/N of the remaining individuals in the
population are put in the second bucket, plus any individuals whose
fitness equals the fittest individual in the second bucket. And so on. 
This continues until we've run out of individuals to put into buckets. 
The idea is to make sure that individuals with the same fitness are all
placed into the same bucket.  The "fitness" of an individual, for
purposes of lexicographic selection, is now his bucket number.

We did not find a direct bucketing number-of-buckets parameter which was
good across several problem domains.  We found 100 was good for
artificial ant, 250 for 11-bit Multiplexer and 5-bit Parity, and 25 for
Symbolic Regression.  You'll need to experiment a bit.  Here's the
settings for Multiplexer:

[BASE] = ec.parsimony.BucketTournamentSelection
# The size of the tournament
[BASE].size = 7
# The number of buckets
[BASE].num-buckets = 250

The default base is
  select.bucket-tournament



RATIO BUCKETED TOURNAMENT SELECTION

Ratio Bucketing improves a bit over direct bucketing.  Here, the idea is
to push low-fitness individuals in to large buckets and place
high-fitness individuals into smaller buckets, even as small as the
individual itself.  This allows more fitness-based distinction among the
"important" individuals in the search (the fitter ones) and puts more
parsimony pressure in the "less important" individuals.  We do this by
defining a ratio of remaining individuals in the population to put in
the next bucket.  Let's say this ratio is 1/R.  We put the 1/R worst
individuals of the population in lowest bucket, plus all remaining
individuals in the population whose fitness is equal to the fittest
individual in the bucket.  We then put the 1/4 next worst remaining
individuals in the next bucket, plus all remaining individuals in the
population whose fitness is equal to the fittest individual in the
second bucket.  And so on, until all individuals have been placed into
buckets.  The "fitness" of an individual, for purposes of lexicographic
selection, is now his bucket number.

Like direct bucketing, we did not find a value of R which was good
across several problem domains.  2 was good for artificial ant, 11-bit
mutiplexer, and regression.  but 6 ws good for 5-bit parity.  So you'll
need to experiment.  Here's the settings for Multiplexer:

[BASE] = ec.parsimony.RatioBucketTournamentSelection
# The size of the tournament
[BASE].size = 7
# The number of buckets
[BASE].ratio = 2

The default base is
        select.ratio-bucket-tournament



TARPEIAN SELECTION

Tarpeian is fairly simple but clever.  PRIOR to evaluation, we sort the
population by size, then identify the M individuals which have
above-average size. From those M individuals we "kill" some N percent. 
Notice that M may vary from population to population depending on the
variance of size among the individuals.

By "kill" we mean that we set the fitness of those individuals to a very
bad value, and also mark them as evaluated so the Evaluator doesn't
bother evaluating them.

Not evaluating the individuals is really important: if every individual
is evaluated, Tarpeian is actually pretty costly compared to other
methods.  But if we prematurely "kill" the individuals, then Tarpeian is
pretty competitive if you count total number of evaluations.

Because Tarpeian must do its work prior to evaluation, it can't operate
as a selection operator in ECJ's framework.  Instead, we've arranged for
Tarpeian to be a Statistics subclass which plugs into the
preEvaluationStatistics hook.  To use it, you just hang it off of your
statistics chain.  Assuming you only have one existing Statistics
object, here's how you'd add Tarpeian in a manner which has proven good
across several problem domains:

stat.num-children = 1
stat.child.0 = ec.parsimony.TarpeianStatistics
stat.child.0.kill-proportion = 0.3

Note that our implementation of Tarpeian will operate over *all* of your
subpopuations, even if you don't want that.  You may need to hack it to
operate differently if you have more than one subpopulation and don't
want Tarpeian parsimony on one or more of them.