def choose_babbling_module(self):
if self.model_babbling == "random":
mode = "random"
elif self.model_babbling == "active":
mode = "prop"
interests = {}
for mid in self.modules.keys():
interests[mid] = self.modules[mid].interest()
if mode == 'random':
mid = np.random.choice(interests.keys())
elif mode == 'greedy':
if np.random.random() < self.choice_eps:
mid = np.random.choice(interests.keys())
else:
mid = max(interests, key=interests.get)
elif mode == 'softmax':
temperature = self.choice_eps
w = interests.values()
mid = self.modules.keys()[softmax_choice(w, temperature)]
elif mode == 'prop':
w = interests.values()
mid = self.modules.keys()[prop_choice(w, eps=self.choice_eps)]
self.chosen_modules.append(mid)
return mid
def choose_babbling_module(self, mode='prop'):
interests = {}
for mid in self.modules.keys():
if not ((not self.enable_hand) and mid == "mod1"):
interests[mid] = self.modules[mid].interest()
else:
interests[mid] = 0.
if mode == 'random':
mid = np.random.choice(self.interests.keys())
elif mode == 'greedy':
eps = 0.2
if np.random.random() < eps:
mid = np.random.choice(self.interests.keys())
else:
mid = max(interests, key=interests.get)
elif mode == 'softmax':
temperature = 0.1
w = interests.values()
mid = self.modules.keys()[softmax_choice(w, temperature)]
elif mode == 'prop':
w = interests.values()
mid = self.modules.keys()[prop_choice(w, eps=self.choice_eps)]
self.chosen_modules.append(mid)
return mid
def choose_babbling_module(self, auto_create=False, progress_threshold=1e-2, mode='softmax', weight_by_level=False):
interests = {}
for mid in self.modules.keys():
interests[mid] = self.modules[mid].interest()
self.emit('interest_' + mid, [self.t, interests[mid]])
self.emit('competence_' + mid, [self.t, self.modules[mid].competence()])
max_progress = max(interests.values())
# self.emit('babbling_module', "mod2")
# return "mod2"
#print "max_progress", max_progress
if not auto_create or max_progress > progress_threshold:
if mode == 'random':
mid = np.random.choice(self.modules.keys())
elif mode == 'greedy':
eps = 0.1
if np.random.random() < eps:
mid = np.random.choice(self.modules.keys())
else:
mid = max(interests, key=interests.get)
elif mode == 'softmax':
temperature = 0.1
mids = interests.keys()
w = interests.values()
#print "progresses", w
#print "competences", [mod.competence() for mod in self.modules.values()]
if weight_by_level:
levels = self.hierarchy.module_levels()
for i in range(len(mids)):
f = 2.0
w[i] = w[i] * np.power(f, max(levels.values()) - levels[mids[i]])
#print w
mid = mids[softmax_choice(w, temperature)]
elif mode == 'prop':
mids = interests.keys()
w = interests.values()
#print "progresses", w
#print "competences", [mod.competence() for mod in self.modules.values()]
if weight_by_level:
levels = self.hierarchy.module_levels()
for i in range(len(mids)):
f = 10.0
w[i] = w[i] * np.power(f, max(levels.values()) - levels[mids[i]])
#print w
mid = mids[prop_choice(w, eps=0.1)]
self.chosen_modules[mid] = self.chosen_modules[mid] + 1
#print self.chosen_modules
self.emit('babbling_module', mid)
return mid
else:
return self.create_module()
def choose_interesting_space(self, mode='softmax'):
s_spaces = self.config.s_spaces
interests = {}
for s_space in s_spaces.keys():
if s_space == "s_h":
interests[s_space] = self.modules["mod1"].interest()
elif s_space == "s_t1":
interests[s_space] = self.modules["mod2"].interest()
# elif s_space == "s_t2":
# interests[s_space] = self.modules["mod3"].interest()
elif s_space == "s_o":
interests[s_space] = np.sum([self.modules[mid].interest() for mid in ["mod3", "mod4"]])
self.emit('interests', [self.t, interests])
if mode == 'random':
s_space = np.random.choice(self.interests.keys())
elif mode == 'greedy':
eps = 0.2
if np.random.random() < eps:
s_space = np.random.choice(self.interests.keys())
else:
s_space = max(interests, key=interests.get)
elif mode == 'softmax':
temperature = 0.1
w = interests.values()
s_space = s_spaces.keys()[softmax_choice(w, temperature)]
elif mode == 'prop':
w = interests.values()
s_space = s_spaces.keys()[prop_choice(w, eps=0.2)]
if self.t % 200 == 1:
print
print 'iterations', self.t - 1
print "competences", np.array([self.modules[mid].competence() for mid in self.modules.keys()])
print "progresses", np.array([self.modules[mid].interest_model.current_progress for mid in self.modules.keys()])
print "interests", np.array([self.modules[mid].interest() for mid in self.modules.keys()])
print "sm db n points", [len(self.modules[mid].sensorimotor_model.model.imodel.fmodel.dataset) for mid in self.modules.keys()]
print "im db n points", [len(self.modules[mid].interest_model.data_xc) for mid in self.modules.keys()]
print self.chosen_modules
print "made tool moved object", self.credit_tool_move
print "made hand moved object", self.credit_hand_move
self.chosen_spaces[s_space] = self.chosen_spaces[s_space] + 1
return s_space
def choose_babbling_module(self):
interests = {}
for mid in self.modules.keys():
interests[mid] = self.modules[mid].interest()
if self.model_babbling == 'random':
#mid = np.random.choice(interests.keys())
if np.random.random() < self.arm_goal_selection:
mid = np.random.choice(self.arm_modules)
else:
mid = np.random.choice(self.diva_modules)
elif self.model_babbling == 'hand_object_sound':
if np.random.random() < 1. / 3.:
mid = 'mod1'
elif np.random.random() < 1. / 2.:
mid = np.random.choice([
'mod2', 'mod3', 'mod4', 'mod5', 'mod10', 'mod11', 'mod12'
])
else:
mid = np.random.choice(['mod6', 'mod13', 'mod14'])
elif self.model_babbling == 'object_sound':
if np.random.random() < 1. / 2.:
mid = np.random.choice([
'mod1', 'mod2', 'mod3', 'mod4', 'mod5', 'mod10', 'mod11',
'mod12'
])
else:
mid = np.random.choice(['mod6', 'mod13', 'mod14'])
elif self.model_babbling == 'greedy':
if np.random.random() < self.choice_eps:
mid = np.random.choice(interests.keys())
else:
mid = max(interests, key=interests.get)
elif self.model_babbling == 'softmax':
temperature = self.choice_eps
w = interests.values()
mid = self.modules.keys()[softmax_choice(w, temperature)]
elif self.model_babbling == 'prop':
w = interests.values()
mid = self.modules.keys()[prop_choice(w, eps=self.choice_eps)]
self.chosen_modules.append(int(mid[3:]))
return mid
def choose_babbling_module(self, auto_create=False, progress_threshold=1e-2, mode="softmax", weight_by_level=False):
interests = {}
for mid in self.modules.keys():
interests[mid] = self.modules[mid].interest()
self.emit("interests", [self.t, interests])
if mode == "random":
mid = np.random.choice(self.modules.keys())
elif mode == "greedy":
eps = 0.1
if np.random.random() < eps:
mid = np.random.choice(self.modules.keys())
else:
mid = max(interests, key=interests.get)
elif mode == "softmax":
temperature = 0.1
mids = interests.keys()
w = interests.values()
if weight_by_level:
levels = self.hierarchy.module_levels()
for i in range(len(mids)):
f = 2.0
w[i] = w[i] * np.power(f, max(levels.values()) - levels[mids[i]])
# print w
mid = mids[softmax_choice(w, temperature)]
elif mode == "prop":
mids = interests.keys()
w = interests.values()
# if self.t % 100 == 0:
# print
# print 'iteration', self.t
# print "competences", np.array([self.modules[mid].competence() for mid in self.modules.keys()])
# if sum(w) > 0:
# print "interests", np.array(w)
# print "interesting objects:", int(((w[3] + w[6]) / sum(w))*100), "%"
# print "cumulated:", int((self.chosen_modules["mod4"] + self.chosen_modules["mod7"]) / float(sum(self.chosen_modules.values())) * 100), "%"
# print self.chosen_modules
# print "competences", [mod.competence() for mod in self.modules.values()]
if weight_by_level:
levels = self.hierarchy.module_levels()
for i in range(len(mids)):
f = 10.0
w[i] = w[i] * np.power(f, max(levels.values()) - levels[mids[i]])
# print w
mid = mids[prop_choice(w, eps=0.1)]
# print
# print
# print self.chosen_modules
# print "chosen module:", mid
elif mode == "prop-min":
mids = interests.keys()
w = interests.values()
# if self.t % 100 == 0:
# print "interests", np.array(w), 'iteration', self.t, "competences", np.array([self.modules[mid].competence() for mid in self.modules.keys()]), int(((w[3] + w[4]-2*min(w)) / sum(np.array(w)-min(w)))*100), "%"
mid = mids[prop_choice(np.array(w) - min(w), eps=0.1)]
self.chosen_modules[mid] = self.chosen_modules[mid] + 1
# print self.chosen_modules
self.emit("babbling_module", mid)
return mid
def choose_babbling_module(self,
auto_create=False,
progress_threshold=1e-2,
mode='softmax',
weight_by_level=False):
interests = {}
for mid in self.modules.keys():
interests[mid] = self.modules[mid].interest()
self.emit('interests', [self.t, interests])
if mode == 'random':
mid = np.random.choice(self.modules.keys())
elif mode == 'greedy':
eps = 0.1
if np.random.random() < eps:
mid = np.random.choice(self.modules.keys())
else:
mid = max(interests, key=interests.get)
elif mode == 'softmax':
temperature = 0.1
mids = interests.keys()
w = interests.values()
if weight_by_level:
levels = self.hierarchy.module_levels()
for i in range(len(mids)):
f = 2.0
w[i] = w[i] * np.power(
f,
max(levels.values()) - levels[mids[i]])
#print w
mid = mids[softmax_choice(w, temperature)]
elif mode == 'prop':
mids = interests.keys()
w = interests.values()
# if self.t % 100 == 0:
# print
# print 'iteration', self.t
# print "competences", np.array([self.modules[mid].competence() for mid in self.modules.keys()])
# if sum(w) > 0:
# print "interests", np.array(w)
# print "interesting objects:", int(((w[3] + w[6]) / sum(w))*100), "%"
# print "cumulated:", int((self.chosen_modules["mod4"] + self.chosen_modules["mod7"]) / float(sum(self.chosen_modules.values())) * 100), "%"
# print self.chosen_modules
#print "competences", [mod.competence() for mod in self.modules.values()]
if weight_by_level:
levels = self.hierarchy.module_levels()
for i in range(len(mids)):
f = 10.0
w[i] = w[i] * np.power(
f,
max(levels.values()) - levels[mids[i]])
#print w
mid = mids[prop_choice(w, eps=0.1)]
# print
# print
# print self.chosen_modules
# print "chosen module:", mid
elif mode == 'prop-min':
mids = interests.keys()
w = interests.values()
# if self.t % 100 == 0:
# print "interests", np.array(w), 'iteration', self.t, "competences", np.array([self.modules[mid].competence() for mid in self.modules.keys()]), int(((w[3] + w[4]-2*min(w)) / sum(np.array(w)-min(w)))*100), "%"
mid = mids[prop_choice(np.array(w) - min(w), eps=0.1)]
self.chosen_modules[mid] = self.chosen_modules[mid] + 1
#print self.chosen_modules
self.emit('babbling_module', mid)
return mid
