def predict_smart(self, x_test):
y = []
number_of_test_examples = len(x_test)
for i in xrange(number_of_test_examples):
search_tree = SearchTree(self.number_of_labels)
root = search_tree.add_vertex("root_", 0, 1.0, 1.0, [], x_test[i])
h = []
heapq.heappush(h, (root.get_inversed_cond_prob(), root))
best_leaf = None
best_score = 0.0
while len(h) != 0:
vertex = heapq.heappop(h)[1]
if vertex.get_conditional_prob() < best_score:
break
else:
greedy_leaf = self.look_ahead(search_tree, vertex, h, best_score)
if greedy_leaf is None:
continue
else:
best_leaf = greedy_leaf
best_score = greedy_leaf.get_conditional_prob()
y.append(best_leaf.get_labels())
return np.array(y)
def predict_smart(self, x_test):
y = []
number_of_test_examples = len(x_test)
for i in xrange(number_of_test_examples):
search_tree = SearchTree(self.number_of_labels)
root = search_tree.add_vertex("root_", 0, 1.0, 1.0, [], x_test[i])
h = []
heapq.heappush(h, (root.get_inversed_cond_prob(), root))
best_leaf = None
best_score = 0.0
while len(h) != 0:
vertex = heapq.heappop(h)[1]
if vertex.get_conditional_prob() < best_score:
break
else:
greedy_leaf = self.look_ahead(search_tree, vertex, h,
best_score)
if greedy_leaf is None:
continue
else:
best_leaf = greedy_leaf
best_score = greedy_leaf.get_conditional_prob()
y.append(best_leaf.get_labels())
return np.array(y)
def predict_full(self, x_test):
y = []
number_of_test_examples = len(x_test)
for i in xrange(number_of_test_examples):
search_tree = SearchTree(self.number_of_labels)
root = search_tree.add_vertex("root_", 0, 1.0, 1.0, [], x_test[i])
self.generate_search_tree(search_tree, root)
best_leaf = search_tree.find_best_leaf()
y.append(best_leaf.get_labels())
# print "Done: " + str(i) + "/" + str(number_of_training_examples)
return np.array(y)
def predict_full(self, x_test):
y = []
number_of_test_examples = len(x_test)
for i in xrange(number_of_test_examples):
search_tree = SearchTree(self.number_of_labels)
root = search_tree.add_vertex("root_", 0, 1.0, 1.0, [], x_test[i])
self.generate_search_tree(search_tree, root)
best_leaf = search_tree.find_best_leaf()
y.append(best_leaf.get_labels())
# print "Done: " + str(i) + "/" + str(number_of_training_examples)
return np.array(y)
def predict_ucs(self, x_test, epsilon):
y = []
number_of_test_examples = len(x_test)
for i in xrange(number_of_test_examples):
search_tree = SearchTree(self.number_of_labels)
root = search_tree.add_vertex("root_", 0, 1.0, 1.0, [], x_test[i])
q = []
greedy_q = []
heapq.heappush(q, (root.get_inversed_cond_prob(), root))
best_leaf = None
while len(q) != 0:
vertex = heapq.heappop(q)[1]
if len(vertex.get_labels()
) == self.number_of_labels: # If leaf
best_leaf = vertex
break
else:
children = self.generate_children(search_tree, vertex)
no_children_inserted = True
for child in children:
# print child.get_inversed_cond_prob()
if child.get_probability() > epsilon:
heapq.heappush(
q, (child.get_inversed_cond_prob(), child))
no_children_inserted = False
if no_children_inserted:
heapq.heappush(
greedy_q,
(vertex.get_inversed_cond_prob(), vertex))
if best_leaf is None:
best_score = 0.0
while len(greedy_q) != 0:
vertex = heapq.heappop(greedy_q)[1]
greedy_leaf = self.look_ahead(search_tree, vertex, [],
best_score)
if greedy_leaf is None:
continue
else:
best_leaf = greedy_leaf
best_score = greedy_leaf.get_conditional_prob()
# print best_leaf, best_score
# print("------")
y.append(best_leaf.get_labels())
# print "Done: " + str(i) + "/" + str(number_of_training_examples)
return np.array(y)
def predict_ucs(self, x_test, epsilon):
y = []
number_of_test_examples = len(x_test)
for i in xrange(number_of_test_examples):
search_tree = SearchTree(self.number_of_labels)
root = search_tree.add_vertex("root_", 0, 1.0, 1.0, [], x_test[i])
q = []
greedy_q = []
heapq.heappush(q, (root.get_inversed_cond_prob(), root))
best_leaf = None
while len(q) != 0:
vertex = heapq.heappop(q)[1]
if len(vertex.get_labels()) == self.number_of_labels: # If leaf
best_leaf = vertex
break
else:
children = self.generate_children(search_tree, vertex)
no_children_inserted = True
for child in children:
# print child.get_inversed_cond_prob()
if child.get_probability() > epsilon:
heapq.heappush(q, (child.get_inversed_cond_prob(), child))
no_children_inserted = False
if no_children_inserted:
heapq.heappush(greedy_q, (vertex.get_inversed_cond_prob(), vertex))
if best_leaf is None:
best_score = 0.0
while len(greedy_q) != 0:
vertex = heapq.heappop(greedy_q)[1]
greedy_leaf = self.look_ahead(search_tree, vertex, [], best_score)
if greedy_leaf is None:
continue
else:
best_leaf = greedy_leaf
best_score = greedy_leaf.get_conditional_prob()
# print best_leaf, best_score
# print("------")
y.append(best_leaf.get_labels())
# print "Done: " + str(i) + "/" + str(number_of_training_examples)
return np.array(y)
def offset(mapA, mapB):
#print('query start')
init = time.time() * 1000
mean_dist = 0
st = SearchTree()
#print('sizes:',len(mapA), len(mapB))
random.shuffle(mapA) #Reduce tree height
for i in mapA:
#print(i[0], i[1])
st.push(i[0], i[1])
#print('building done. Time was', (time.time() * 1000 - init))
init = time.time() * 1000
for j in mapB:
#print(j, st.query(j[0], j[1], False))
res = st.query(j[0], j[1], False)
mean_dist += res[0]
#print('query end. Time was', (time.time() * 1000 - init), 'Visited count:', res[1])
return (mean_dist / len(mapB))
def __init__(self,
world,
name=None,
init_vertex=1,
bonus_vandal_records=None):
super(SmartAStar,
self).__init__(world=world,
name=name,
init_vertex=init_vertex,
bonus_vandal_records=bonus_vandal_records)
self.search_tree = SearchTree(init_state=self.init_state,
strategy='A*',
vandal_records=bonus_vandal_records)
def __init__(self,
world,
expand_limit,
name=None,
init_vertex=1,
bonus_vandal_records=None):
super(SmartRTA,
self).__init__(world=world,
name=name,
init_vertex=init_vertex,
bonus_vandal_records=bonus_vandal_records)
self.search_tree = SearchTree(init_state=self.init_state,
strategy='RTA',
expand_limit=expand_limit,
vandal_records=bonus_vandal_records)
output_size=board_size * board_size,
hidden_layers=config.HIDDEN_LAYERS,
activation=config.ACTIVATION).float()
anet = ANET_MODEL(anet_nn,
max_cases_in_buffer=config.REPLAY_BUFFER_MAX_SIZE)
optimizer = config.OPTIMIZERS[config.OPTIMIZER](anet_nn.parameters(),
lr=lr)
for episode in range(num_episodes):
while state_manager.game_is_on():
random_number = random.uniform(0, 1)
if random_number < 0.3:
action = state_manager.get_random_action()
else:
search_tree = SearchTree(state_manager.get_current_state(),
exploration_bonus)
action = search_tree.simulate_games_and_find_action(
num_simulations, state_manager, anet)
state_manager.take_action(action)
# --------- LOGGING THE TIME MODELS USE TO RAIN --------------
time_finished = time.time()
time_passed = str(round((time_finished - time_start) / 60, 2))
log_string = l = str(episode) + " of " + str(
num_episodes) + " finished. TIME PASSED (minutes): " + time_passed
utils.logging(folder_results, log_string)
# ------------------------------------------------------------
anet.train_model(optimizer,
minibatch_size=config.REPLAY_BUFFER_MINIBATCH_SIZE)
