def testKnownTrees(self):
print "\n\n----- calculating fitness of known trees -----"
# create trees - MI, OMES, whatever - that have a special form and
# check their fitness
treenames = ['MI','OMES']
for t in treenames:
tree = CGAGenerator.generate_special_tree(t)
print '%s : %s' % (t,tree.getString())
print 'Weighted accuracy fitness : ',self.mySimulation.evaluate_fitness_weighted_accuracy(tree)
print 'Top N weighted accuracy fitness : ',self.mySimulation.evaluate_fitness_topN_weighted_accuracy(tree)
print 'Distance matrix fitness : ',self.mySimulation.evaluate_fitness_distance_matrix(tree)
def initialize_tree(self):
"""Creates a single random tree using the parameters in self.treeGenDict. This functionality
has been separated from self.populate() in order to make headless chicken crossover (in which we
crossover with a randomly generated tree) easier."""
r = self.treeGenDict['r']
if self.treeGenDict['treetype'] == 'exponential':
if self.treeGenDict['p'] < 1.0:
p = self.treeGenDict['p']
else:
# ensures consistent parameter values
p = 1/self.treeGenDict['p']
return CGAGenerator.expgenerate(p,r)
elif self.treeGenDict['treetype'] == 'fixed':
if self.treeGenDict['p'] > 1:
treeSize = self.treeGenDict['p']
else:
treeSize = MATH.int(1.0/self.treeGenDict['p'])
return CGAGenerator.generate(treeSize,r)
else:
raise TypeError, 'Unknown tree type %s' % self.treeGenDict['treetype']
def mate(self, parentOne, parentTwo):
"""Accepts two parents and returns ONE offspring; if the offspring is unchanged from one of
the parents, the fitness values are not re-evaluated, just copied. Only one category of
mutation is performed on a single offspring."""
# one parent will form the basis for the new offspring; the other is just there for
# potential crossover (by default the mother will become the offspring)
mother, father = parentOne.copy(), parentTwo.copy()
# fitEval will be set to True if any mutations occur that make the offspring different from
# its copied parent; this way we avoid unnecessary fitness evaluations
fitEval = False
# only one category of mutation is allowed per mating event
r = urand()
if r < self.pC: # parental crossover
fitEval = True
mNode = rchoice(mother.tree.getNodes())
fNode = rchoice(father.tree.getNodes())
CGAGenerator.single_crossover(mNode,fNode)
elif r < self.pC + self.pHC: # headless chicken crossover
fitEval = True
mNode = rchoice(mother.tree.getNodes())
rNode = rchoice(self.initialize_tree().getNodes())
CGAGenerator.single_crossover(mNode,rNode)
elif r < self.pC + self.pHC + self.pM: # point mutation (uses pM/node for mutation prob.)
fitEval = True
for n in mother.tree.getNodes():
if urand() < self.pM:
CGAGenerator.point_mutate(mother.tree, n)
fitEval = True
elif r < self.pC + self.pHC + self.pM + self.pP: # pruning - guaranteed to do one pruning operation
fitEval = True
mNode = rchoice(mother.tree.getNodes())
CGAGenerator.prune(mother.tree,mNode)
elif r < self.pC + self.pHC + self.pM + self.pP + self.pG: # growth - guaranteed to do one growth op.
fitEval = True
mTerm = rchoice(mother.tree.getTermini())
CGAGenerator.grow(mother.tree,mTerm)
else: # offspring will just be a copy
pass
if fitEval:
mother.fitness,mother.parsimony,mother.finitewts = self.evaluate_fitness(mother.tree)
return mother
