def test_add_learners(self):
team = Team(dummy_init_params)
original_learners = team.learners
l1 = create_dummy_learner()
l2 = create_dummy_learner()
l1_original_team_references = l1.numTeamsReferencing()
l2_original_team_references = l2.numTeamsReferencing()
team.addLearner(l1)
team.addLearner(l2)
#Ensure learner ref counts incremented
self.assertEqual(l1_original_team_references + 1,
l1.numTeamsReferencing())
self.assertEqual(l2_original_team_references + 1,
l2.numTeamsReferencing())
#Ensure learner list grew
self.assertEqual(2, len(team.learners))
#Ensure no incomming learners appear in the team's inlearner list
self.assertEqual(0, len(team.inLearners))
def generate(self):
oLearners = list(self.learners)
oTeams = list(self.teams)
# multiActs for action pool for multiaction mutation
if self.multiAction:
multiActs = []
for learner in oLearners:
if learner.isActionAtomic():
multiActs.append(list(learner.action))
else:
multiActs = None
while (len(self.teams) < self.teamPopSize
or self.countRootTeams() < self.rTeamPopSize):
# get parent root team, and child to be based on that
parent = random.choice(self.rootTeams)
child = Team()
# child starts just like parent
for learner in parent.learners:
child.addLearner(learner)
if self.uniqueProgThresh > 0:
inputs, outputs = self.getLearnersInsOuts(oLearners)
else:
inputs = None
outputs = None
# then mutates
child.mutate(self.pDelLrn,
self.pAddLrn,
self.pMutLrn,
oLearners,
self.pMutProg,
self.pMutAct,
self.pActAtom,
self.actions,
oTeams,
self.pDelInst,
self.pAddInst,
self.pSwpInst,
self.pMutInst,
multiActs,
self.pSwapMultiAct,
self.pChangeMultiAct,
self.uniqueProgThresh,
inputs=inputs,
outputs=outputs,
update=True)
self.teams.append(child)
self.rootTeams.append(child)
def create_dummy_team(num_learners=2):
team = Team(dummy_init_params)
learners = []
for x in range(0, num_learners):
learner = create_dummy_learner()
learners.append(learner)
team.addLearner(learner)
return team, learners
def initializePopulations(self):
for i in range(self.teamPopSize):
# create 2 unique actions and learners
a1,a2 = random.sample(range(len(self.actionCodes)), 2)
l1 = Learner(self.mutateParams,
program=Program(maxProgramLength=self.initMaxProgSize,
nOperations=self.nOperations,
nDestinations=self.nRegisters,
inputSize=self.inputSize,
initParams=self.mutateParams),
actionObj=ActionObject(action=a1, initParams=self.mutateParams),
numRegisters=self.nRegisters)
l2 = Learner(self.mutateParams,
program=Program(maxProgramLength=self.initMaxProgSize,
nOperations=self.nOperations,
nDestinations=self.nRegisters,
inputSize=self.inputSize,
initParams=self.mutateParams),
actionObj=ActionObject(action=a2, initParams=self.mutateParams),
numRegisters=self.nRegisters)
# save learner population
self.learners.append(l1)
self.learners.append(l2)
# create team and add initial learners
team = Team(initParams=self.mutateParams)
team.addLearner(l1)
team.addLearner(l2)
# add more learners
moreLearners = random.randint(0, self.initMaxTeamSize-2)
for i in range(moreLearners):
# select action
act = random.choice(range(len(self.actionCodes)))
# create new learner
learner = Learner(initParams=self.mutateParams,
program=Program(maxProgramLength=self.initMaxProgSize,
nOperations=self.nOperations,
nDestinations=self.nRegisters,
inputSize=self.inputSize,
initParams=self.mutateParams),
actionObj=ActionObject(action=act, initParams=self.mutateParams),
numRegisters=self.nRegisters)
team.addLearner(learner)
self.learners.append(learner)
# save to team populations
self.teams.append(team)
self.rootTeams.append(team)
def test_no_duplicate_learners(self):
program = create_dummy_program()
l1 = create_dummy_learner()
l2 = create_dummy_learner()
# Set the program to be the same for both learners
l1.program = copy.deepcopy(program)
l2.program = copy.deepcopy(program)
# Create a team
team = Team(dummy_init_params)
# Add the first learner
team.addLearner(l1)
def initializePopulations(self, initMaxTeamSize, initMaxProgSize,
registerSize):
for i in range(self.teamPopSize):
# create 2 unique actions and learners
if self.multiAction == False:
a1, a2 = random.sample(self.actions, 2)
else:
a1 = [random.uniform(0, 1) for _ in range(self.actions)]
a2 = [random.uniform(0, 1) for _ in range(self.actions)]
l1 = Learner(program=Program(maxProgramLength=initMaxProgSize),
action=a1,
numRegisters=registerSize)
l2 = Learner(program=Program(maxProgramLength=initMaxProgSize),
action=a2,
numRegisters=registerSize)
# save learner population
self.learners.append(l1)
self.learners.append(l2)
# create team and add initial learners
team = Team()
team.addLearner(l1)
team.addLearner(l2)
# add more learners
moreLearners = random.randint(0, initMaxTeamSize - 2)
for i in range(moreLearners):
# select action
if self.multiAction == False:
act = random.choice(self.actions)
else:
act = [random.uniform(0, 1) for _ in range(self.actions)]
# create new learner
learner = Learner(
program=Program(maxProgramLength=initMaxProgSize),
action=act,
numRegisters=registerSize)
team.addLearner(learner)
self.learners.append(learner)
# save to team populations
self.teams.append(team)
self.rootTeams.append(team)
def generate(self, extraTeams=None):
# extras who are already part of the team population
protectedExtras = []
extrasAdded = 0
# add extras into the population
if extraTeams is not None:
for team in extraTeams:
if team not in self.teams:
self.teams.append(team)
extrasAdded += 1
else:
protectedExtras.append(team)
oLearners = list(self.learners)
oTeams = list(self.teams)
# update generation in mutateParams
self.mutateParams["generation"] = self.generation
# get all the current root teams to be parents
while (len(self.teams) < self.teamPopSize + extrasAdded or
(self.rootBasedPop and self.countRootTeams() < self.teamPopSize)):
# get parent root team, and child to be based on that
parent = random.choice(self.rootTeams)
child = Team(initParams=self.mutateParams)
# child starts just like parent
for learner in parent.learners:
child.addLearner(learner)
# then mutates
child.mutate(self.mutateParams, oLearners, oTeams)
self.teams.append(child)
# remove unused extras
if extraTeams is not None:
for team in extraTeams:
if team.numLearnersReferencing() == 0 and team not in protectedExtras:
self.teams.remove(team)
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)