def choose_actions(self, subject, actioni):
action_values={}
for a in world.possible_actions():
IG=self.alldata[subject][actioni]['ActionValues'][a][0]
PG=self.alldata[subject][actioni]['ActionValues'][a][1]
#PGv=1-PG #invert, IG is final entropy, minimized! PG=0 -> PGv=1; PG=1 -> PGv=0
#action_values[a]=self.theta*IG+(1-self.theta)*PGv
action_values[a]=-self.theta*IG-(1-self.theta)*PG
#min_value=action_values[min(action_values, key=lambda x: x[1])] #take the min value --WRONG!!!
min_value=min(action_values.itervalues())
#min_actions=[a for a in action_values.keys() if action_values[a]==min_value] #find ALL actions that achieve it --should be ok, cleaner below
min_actions=[a for a,v in action_values.iteritems() if v==min_value]
#discard repeated action choice
if actioni>0:
prev_action=self.alldata[subject][actioni-1]['SubjectAction']
else:
prev_action=()
if len(set(min_actions)-set([prev_action])) == 0:
del action_values[prev_action]
min_value=min(action_values.itervalues())
min_actions=[a for a,v in action_values.iteritems() if v==min_value]
else:
min_actions=set(min_actions)-set([prev_action])
return list(min_actions)
def choose_actions(self, subject, actioni):
action_values = {}
for a in world.possible_actions():
IG = self.alldata[subject][actioni]['ActionValues'][a][0]
PG = self.alldata[subject][actioni]['ActionValues'][a][1]
#PGv=1-PG #invert, IG is final entropy, minimized! PG=0 -> PGv=1; PG=1 -> PGv=0
#action_values[a]=self.theta*IG+(1-self.theta)*PGv
action_values[a] = -self.theta * IG - (1 - self.theta) * PG
#min_value=action_values[min(action_values, key=lambda x: x[1])] #take the min value --WRONG!!!
min_value = min(action_values.itervalues())
#min_actions=[a for a in action_values.keys() if action_values[a]==min_value] #find ALL actions that achieve it --should be ok, cleaner below
min_actions = [
a for a, v in action_values.iteritems() if v == min_value
]
#discard repeated action choice
if actioni > 0:
prev_action = self.alldata[subject][actioni - 1]['SubjectAction']
else:
prev_action = ()
if len(set(min_actions) - set([prev_action])) == 0:
del action_values[prev_action]
min_value = min(action_values.itervalues())
min_actions = [
a for a, v in action_values.iteritems() if v == min_value
]
else:
min_actions = set(min_actions) - set([prev_action])
return list(min_actions)
def choose_action(self, prev_data=None): mingain=1000 astars=[] for a in world.possible_actions(): this_gain=self.entropy_gain(a, prev_data) #print a, this_gain if this_gain <= mingain: astars.append(a) mingain=this_gain return random.choice(astars)
def choose_action(self, prev_data=None):
mingain = 1000
astars = []
for a in world.possible_actions():
this_gain = self.entropy_gain(a, prev_data)
#print a, this_gain
if this_gain <= mingain:
astars.append(a)
mingain = this_gain
return random.choice(astars)
def choose_actions(self, prev_data=[]):
"""
Same as choose actions, but return all equivalents
"""
mingain = 1000
astars = []
for a in world.possible_actions():
if len(prev_data) > 0:
if a == prev_data[-1].action:
continue
this_gain = self.expected_final_entropy(a, prev_data)
if this_gain < mingain:
astars = [a]
mingain = this_gain
elif this_gain == mingain:
astars.append(a)
return astars, mingain
def analyze(self):
data=self.load_data()
subjects=data.get_kids()[:self.N]
for subject in subjects:
print 'Analyzing subject {0}...'.format(subject)
self.alldata[subject]={}#defaultdict(dict)
max_action=min(data.get_kid_nactions(subject),self.A)
subject_sequence=data.data[subject][:max_action]
for actioni in range(max_action):
#print 'Action {0} of {1}'.format(actioni, max_action)
self.alldata[subject][actioni]={}
subject_action=data.data[subject][actioni].get_action()
self.alldata[subject][actioni]['SubjectAction']=subject_action
theory_model=learners.TheoryLearner()
pg_model=learners.ActivePlayer()
# model_actions, model_gain=pg_model.choose_actions(subject_sequence[:actioni])
self.alldata[subject][actioni]['ActionValues']={}
for a in world.possible_actions():
#EIG=theory_model.expected_final_entropy(a, subject_sequence[:actioni])
EIG=-1*theory_model.expected_information_gain(a, subject_sequence[:actioni])
PG=pg_model.success_probability(a, subject_sequence[:actioni])
self.alldata[subject][actioni]['ActionValues'][a]=(EIG,PG)
# theory_model=learners.TheoryLearner()
# model_actions, model_gain=theory_model.choose_actions(subject_sequence[:actioni])
# self.alldata[subject][actioni]['TMA']=model_actions
# self.alldata[subject][actioni]['SEIG']=\
# entropy_gains.theory_expected_final_entropy(subject_action, subject_sequence[:actioni])
# self.alldata[subject][actioni]['TMEIG']=model_gain
# pg_model=learners.ActivePlayer()
# model_actions, model_gain=pg_model.choose_actions(subject_sequence[:actioni])
# self.alldata[subject][actioni]['PMA']=model_actions
# self.alldata[subject][actioni]['PMSP']=model_gain
self.save()
def choose_actions(self, prev_data=[]): """ Same as choose actions, but return all equivalents """ mingain=1000 astars=[] for a in world.possible_actions(): if len(prev_data)>0: if a==prev_data[-1].action: continue this_gain=self.expected_final_entropy(a, prev_data) if this_gain < mingain: astars=[a] mingain=this_gain elif this_gain == mingain: astars.append(a) return astars, mingain
def choose_actions(self, prev_data=[]):
"""
Same as choose actions, but return all equivalents
"""
maxprob = 0
astars = []
for a in world.possible_actions():
if len(prev_data) > 0:
if a == prev_data[-1].action:
continue
this_prob = self.success_probability(a, prev_data)
if this_prob > maxprob:
astars = [a]
maxprob = this_prob
elif this_prob == maxprob:
astars.append(a)
#choice=random.choice(astars)
return astars, maxprob
def choose_actions(self, prev_data=[]): """ Same as choose actions, but return all equivalents """ maxprob=0 astars=[] for a in world.possible_actions(): if len(prev_data)>0: if a==prev_data[-1].action: continue this_prob=self.success_probability(a, prev_data) if this_prob > maxprob: astars=[a] maxprob=this_prob elif this_prob == maxprob: astars.append(a) #choice=random.choice(astars) return astars, maxprob
def choose_action(self, prev_data=[]):
#mingain=1000
maxprob = 0
astars = []
for a in world.possible_actions():
if len(prev_data) > 0:
if a == prev_data[-1].action:
continue
#this_gain=self.expected_final_entropy(a, prev_data)
this_prob = self.success_probability(a, prev_data)
if this_prob > maxprob:
astars = [a]
maxprob = this_prob
#mingain=this_gain
#elif this_gain == mingain:
elif this_prob == maxprob:
astars.append(a)
choice = random.choice(astars)
#self.experience.append(world.make_action(choice))
return choice
def choose_action(self, prev_data=[]): #mingain=1000 maxprob=0 astars=[] for a in world.possible_actions(): if len(prev_data)>0: if a==prev_data[-1].action: continue #this_gain=self.expected_final_entropy(a, prev_data) this_prob=self.success_probability(a, prev_data) if this_prob > maxprob: astars=[a] maxprob=this_prob #mingain=this_gain #elif this_gain == mingain: elif this_prob == maxprob: astars.append(a) choice=random.choice(astars) #self.experience.append(world.make_action(choice)) return choice
def choose_action(self, prev_data=[]): #None):
mingain = 1000
astars = []
for a in world.possible_actions():
if len(prev_data) > 0:
if a == prev_data[-1].action:
continue
this_gain = self.expected_final_entropy(a, prev_data)
if this_gain < mingain:
astars = [a]
mingain = this_gain
elif this_gain == mingain:
astars.append(a)
#while True:
# print astars, prev_data[-1].action
choice = random.choice(astars)
# if len(prev_data)==0:
# break
# elif choice!=prev_data[-1].action:
# break
#self.experience.append(world.make_action(choice))
return choice
def choose_action(self, prev_data=[]):#None): mingain=1000 astars=[] for a in world.possible_actions(): if len(prev_data)>0: if a==prev_data[-1].action: continue this_gain=self.expected_final_entropy(a, prev_data) if this_gain < mingain: astars=[a] mingain=this_gain elif this_gain == mingain: astars.append(a) #while True: # print astars, prev_data[-1].action choice=random.choice(astars) # if len(prev_data)==0: # break # elif choice!=prev_data[-1].action: # break #self.experience.append(world.make_action(choice)) return choice
lposc = [0] * 12
hposc = [0] * 12
lposs = [0] * 12
hposs = [0] * 12
lposi = [0] * 12
hposi = [0] * 12
data = Data.Data()
data.read(astext=False)
dind = data.data[data.get_kids()[3]]
[d.display() for d in dind]
#dind[-1].active=False
print low_model.p_data_action(dind[0], world.possible_actions()[3])
print high_model.p_data_action(dind[0], world.possible_actions()[3])
dind[0].active = True
print low_model.p_data_action(dind[0], world.possible_actions()[3])
print high_model.p_data_action(dind[0], world.possible_actions()[3])
norm = True
lostart = time.clock()
for t in range(12):
lposc[t] = low_model.p_theory_data(t, dcol, normalized=norm)
lposs[t] = low_model.p_theory_data(t, dsha, normalized=norm)
lposi[t] = low_model.p_theory_data(t, dind, normalized=norm)
# l=[]
if acts2[i].active: col=0 if acts2[i].machine[0]==acts2[i].toy[0]: condition[i]='color' else: condition[i]='shape' else: if acts2[i].machine[0]!=acts2[i].toy[0] and acts2[i].machine[1]!=acts2[i].toy[1]: col=2 else: col=1 table[row,col]+=1 import world wacts=world.possible_actions() kacts=np.zeros((6,3),dtype=int) for i in range(30): toy1=acts1[i].toy mac1=acts1[i].machine toy2=acts2[i].toy mac2=acts2[i].machine if acts1[i].active: col=0 if toy1[0]==mac1[0]: prop1='col_p' elif toy1[1]==mac1[1]: prop1='sha_p' else: if toy1[0]==mac1[0]:
today='141119/' #batch='ep-0.05/' batch='varyep/' data_directory=output_directory+today#+batch n_act=2 model='jointfull' filename=model+'-'+str(n_act)+'_tru-20'+'_rreal.txt' datajoi=np.loadtxt(data_directory+filename) print data_directory+filename model='hypfull' filename=model+'-'+str(n_act)+'_tru-20'+'_rreal.txt' datahyp=np.loadtxt(data_directory+filename) print data_directory+filename for i in range(len(datahyp)): print datahyp[i,2]-datahyp[i,0], datajoi[i,2]-datajoi[i,0] import entropy_gains as eg import world pa=world.possible_actions() print eg.hypotheses_expected_final_entropy(pa[0],[])-eg.hypotheses_expected_final_entropy(pa[1],[]) print eg.joint_expected_final_entropy(pa[0],[])-eg.hypotheses_expected_final_entropy(pa[1],[])
import low_model as model import world import entropy_gains as eg for action in world.possible_actions(): print action, eg.hypotheses_expected_final_entropy(action,[])
lposc = [0] * 12
hposc = [0] * 12
lposs = [0] * 12
hposs = [0] * 12
lposi = [0] * 12
hposi = [0] * 12
data = Data.Data()
data.read(astext=False)
dind = data.data[data.get_kids()[3]]
[d.display() for d in dind]
# dind[-1].active=False
print low_model.p_data_action(dind[0], world.possible_actions()[3])
print high_model.p_data_action(dind[0], world.possible_actions()[3])
dind[0].active = True
print low_model.p_data_action(dind[0], world.possible_actions()[3])
print high_model.p_data_action(dind[0], world.possible_actions()[3])
norm = True
lostart = time.clock()
for t in range(12):
lposc[t] = low_model.p_theory_data(t, dcol, normalized=norm)
lposs[t] = low_model.p_theory_data(t, dsha, normalized=norm)
lposi[t] = low_model.p_theory_data(t, dind, normalized=norm)
col = 0
if acts2[i].machine[0] == acts2[i].toy[0]:
condition[i] = 'color'
else:
condition[i] = 'shape'
else:
if acts2[i].machine[0] != acts2[i].toy[0] and acts2[i].machine[
1] != acts2[i].toy[1]:
col = 2
else:
col = 1
table[row, col] += 1
import world
wacts = world.possible_actions()
kacts = np.zeros((6, 3), dtype=int)
for i in range(30):
toy1 = acts1[i].toy
mac1 = acts1[i].machine
toy2 = acts2[i].toy
mac2 = acts2[i].machine
if acts1[i].active:
col = 0
if toy1[0] == mac1[0]:
prop1 = 'col_p'
elif toy1[1] == mac1[1]:
prop1 = 'sha_p'
else:
if toy1[0] == mac1[0]:
