def initPops(self):
# create teams to fill population
for i in range(self.teamPopSize):
# take two distinct initial actions for each of two learners on team
if not self.multiAction: # choose single number
ac1 = self.rand.choice(self.actions)
ac2 = self.rand.choice([a for a in self.actions if a != ac1])
else: # choose list of length self.actions within range
minv = self.actionRange[0]
maxv = self.actionRange[1]
ac1 = [
self.rand.uniform(minv, maxv) for i in range(self.actions)
]
ac2 = [
self.rand.uniform(minv, maxv) for i in range(self.actions)
]
team = Team() # create new team
# add/create first learner
learner = Learner(ac1, self.maxProgramSize, randSeed=self.randSeed)
team.addLearner(learner)
self.learners.append(learner)
# add/create second learner
learner = Learner(ac2, self.maxProgramSize, randSeed=self.randSeed)
team.addLearner(learner)
self.learners.append(learner)
# add other random learners
learnerMax = self.rand.randint(0, self.maxTeamSize - 2)
for i in range(learnerMax):
if not self.multiAction: # choose single number
ac = self.rand.choice(self.actions)
else: # choose list of length self.actions within range
ac = [
self.rand.uniform(minv, maxv)
for i in range(self.actions)
]
learner = Learner(ac,
maxProgSize=self.maxProgramSize,
randSeed=self.randSeed)
team.addLearner(learner)
self.learners.append(learner)
team.uid = TpgTrainer.teamIdCounter
TpgTrainer.teamIdCounter += 1
# add into population
self.teams.append(team)
self.rootTeams.append(team)
def generateNewTeams(self, parents=None):
if parents is None:
parents = list(
self.rootTeams) # parents are all original root teams
# add teams until maxed size
while (len(self.teams) < self.teamPopSize
or (self.rTeamPopSize > 0
and self.getRootTeamsSize() < self.rTeamPopSize)):
# choose 2 random teams as parents
par1 = self.rand.choice(parents)
par2 = self.rand.choice(
[par for par in parents if par is not par1])
# get learners
par1Lrns = set(par1.learners)
par2Lrns = set(par2.learners)
# make 2 children at a time
child1 = Team(birthGen=self.curGen)
child2 = Team(birthGen=self.curGen)
# new children get intersection of parents learners
for learner in par1Lrns.intersection(par2Lrns):
child1.addLearner(learner)
child2.addLearner(learner)
# learners unique to a parent goes to one child or other, with one
# child having higher priority for a learner
for learner in par1Lrns.symmetric_difference(par2Lrns):
superChild = None
subChild = None
if self.rand.choice([True, False]) == True:
superChild = child1
subChild = child2
else:
superChild = child2
subChild = child1
# apply learner to child if can,
# if not, give to other child if can
if (len(superChild.learners) < self.maxTeamSize
and (len(superChild.learners) < 2
or len(subChild.learners) >= 2)):
superChild.addLearner(learner)
else:
subChild.addLearner(learner)
self.mutate(child1) # attempt a mutation
if (set(child1.learners) == set(par1.learners)
or set(child1.learners) == set(par2.learners)):
while not self.mutate(
child1): # attempt mutation untill it works
continue
self.mutate(child2) # attempt a mutation
if (set(child2.learners) == set(par1.learners)
or set(child2.learners) == set(par2.learners)):
while not self.mutate(
child2): # attempt mutation untill it works
continue
child1.uid = TpgTrainer.teamIdCounter
TpgTrainer.teamIdCounter += 1
child2.uid = TpgTrainer.teamIdCounter
TpgTrainer.teamIdCounter += 1
# add children to team populations
self.teams.append(child1)
self.teams.append(child2)
self.rootTeams.append(child1)
self.rootTeams.append(child2)