|
|||||||||
PREV CLASS NEXT CLASS | FRAMES NO FRAMES | ||||||||
SUMMARY: INNER | FIELD | CONSTR | METHOD | DETAIL: FIELD | CONSTR | METHOD |
java.lang.Object | +--ec.BreedingSource | +--ec.BreedingPipeline | +--ec.gp.GPBreedingPipeline | +--ec.gp.breed.MutateDemotePipeline
MutateDemotePipeline works very similarly to the DemoteNode algorithm described in Kumar Chellapilla, "A Preliminary Investigation into Evolving Modular Programs without Subtree Crossover", GP98, and is also similar to the "insertion" operator found in Una-May O'Reilly's thesis, "An Analysis of Genetic Programming".
MutateDemotePipeline tries picks a random tree, then picks randomly from all the demotable nodes in the tree, and demotes one. If its chosen tree has no demotable nodes, or demoting its chosen demotable node would make the tree too deep, it repeats the choose-tree-then-choose-node process. If after tries times it has failed to find a valid tree and demotable node, it gives up and simply copies the individual.
"Demotion" means to take a node n and insert a new node m between n and n's parent. n becomes a child of m; the place where it becomes a child is determined at random from all the type-compatible slots of m. The other child slots of m are filled with randomly-generated terminals. Chellapilla's version of the algorithm always places n in child slot 0 of m. Because this would be unneccessarily restrictive on strong typing, MutateDemotePipeline instead picks the slot at random from all available valid choices.
A "Demotable" node means a node which is capable of demotion given the existing function set. In general to demote a node foo, there must exist in the function set a nonterminal whose return type is type-compatible with the child slot foo holds in its parent; this nonterminal must also have a child slot which is type-compatible with foo's return type.
This method is very expensive in searching nodes for "demotability". However, if the number of types is 1 (the GP run is typeless) then the type-constraint-checking code is bypassed and the method runs a little faster.
Typical Number of Individuals Produced Per produce(...) call
...as many as the source produces
Number of Sources
1
Parameters
base.tries int >= 1 |
(number of times to try finding valid pairs of nodes) |
base.maxdepth int >= 1 |
(maximum valid depth of a mutated tree) |
base.tree.0 0 < int < (num trees in individuals), if exists |
(tree chosen for mutation; if parameter doesn't exist, tree is picked at random) |
Default Base
gp.breed.mutate-demote
Field Summary | |
static int |
NUM_SOURCES
|
static java.lang.String |
P_MAXDEPTH
|
static java.lang.String |
P_MUTATEDEMOTE
|
static java.lang.String |
P_NUM_TRIES
|
Fields inherited from class ec.gp.GPBreedingPipeline |
P_NODESELECTOR, P_TREE, TREE_UNFIXED |
Fields inherited from class ec.BreedingPipeline |
DYNAMIC_SOURCES, mybase, P_NUMSOURCES, P_SOURCE, sources, V_SAME |
Fields inherited from class ec.BreedingSource |
CHECKBOUNDARY, DEFAULT_PRODUCED, NO_PROBABILITY, P_PROB, probability, UNUSED |
Constructor Summary | |
MutateDemotePipeline()
|
Method Summary | |
Parameter |
defaultBase()
Returns the default base for this prototype. |
int |
numSources()
Returns the number of sources to this pipeline. |
int |
produce(int min,
int max,
int start,
int subpopulation,
Individual[] inds,
EvolutionState state,
int thread)
Produces n individuals from the given subpopulation and puts them into inds[start...start+n-1], where n = Min(Max(q,min),max), where q is the "typical" number of individuals the BreedingSource produces in one shot, and returns n. |
java.lang.Object |
protoClone()
Creates a new individual cloned from a prototype, and suitable to begin use in its own evolutionary context. |
void |
setup(EvolutionState state,
Parameter base)
Sets up the BreedingPipeline. |
Methods inherited from class ec.gp.GPBreedingPipeline |
produces |
Methods inherited from class ec.BreedingPipeline |
individualReplaced, maxChildProduction, minChildProduction, preparePipeline, prepareToProduce, sourcesAreProperForm, typicalIndsProduced |
Methods inherited from class ec.BreedingSource |
getProbability, pickRandom, protoCloneSimple, setProbability, setupProbabilities |
Methods inherited from class java.lang.Object |
clone, equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait |
Field Detail |
public static final java.lang.String P_MUTATEDEMOTE
public static final java.lang.String P_NUM_TRIES
public static final java.lang.String P_MAXDEPTH
public static final int NUM_SOURCES
Constructor Detail |
public MutateDemotePipeline()
Method Detail |
public Parameter defaultBase()
Prototype
public int numSources()
BreedingPipeline
numSources
in class BreedingPipeline
public void setup(EvolutionState state, Parameter base)
BreedingSource
The most common modification is to normalize it with some other set of probabilities, then set all of them up in increasing summation; this allows the use of the fast static BreedingSource-picking utility method, BreedingSource.pickRandom(...). In order to use this method, for example, if four breeding source probabilities are {0.3, 0.2, 0.1, 0.4}, then they should get normalized and summed by the outside owners as: {0.3, 0.5, 0.6, 1.0}.
setup
in class BreedingPipeline
ec.BreedingSource
Prototype.setup(EvolutionState,Parameter)
public java.lang.Object protoClone() throws java.lang.CloneNotSupportedException
Prototype
The question here is whether or not this means to perform a "deep" or "light" ("shallow") clone, or something in-between. You may need to deep-clone parts of your object rather than simply copying their references, depending on the situation:
Implementations.
public Object protoClone() throws CloneNotSupportedException
{
return super.clone();
}
public Object protoClone() throws CloneNotSupportedException
{
myobj = (MyObject) (super.clone());
// put your deep-cloning code here...
// ...you should use protoClone and not
// protoCloneSimple to clone subordinate objects...
return myobj;
}
public Object protoClone() throws CloneNotSupportedException
{
MyObject myobj = (MyObject)(super.protoClone());
// put your deep-cloning code here...
// ...you should use protoClone and not
// protoCloneSimple to clone subordinate objects...
return myobj;
}
If you know that your superclasses will never change their protoClone() implementations, you might try inlining them in your overridden protoClone() method. But this is dangerous (though it yields a small net increase).
In general, you want to keep your deep cloning to an absolute minimum, so that you don't have to call protoClone() but one time.
The approach taken here is the fastest that I am aware of while still permitting objects to be specified at runtime from a parameter file. It would be faster to use the "new" operator; but that would require hard-coding that we can't do. Although using java.lang.Object.clone() entails an extra layer that deals with stripping away the "protected" keyword and also wrapping the exception handling (which is a BIG hit, about three times as slow as using "new"), it's still MUCH faster than using java.lang.Class.newInstance(), and also much faster than rolling our own Clone() method.
protoClone
in class BreedingPipeline
public int produce(int min, int max, int start, int subpopulation, Individual[] inds, EvolutionState state, int thread) throws java.lang.CloneNotSupportedException
BreedingSource
produce
in class BreedingSource
|
|||||||||
PREV CLASS NEXT CLASS | FRAMES NO FRAMES | ||||||||
SUMMARY: INNER | FIELD | CONSTR | METHOD | DETAIL: FIELD | CONSTR | METHOD |