def test_predict(self):
"""
Test case for AC-learning and Q-learning predictions
"""
num_actions = 4
def test(input, ct, max):
action_counter = [0] * num_actions
total = 3000
for i in range(total):
actions, states = ct.predict(inputs=input)
assert not states, "states should be empty"
## actions["action"] is a batch of actions
for a in actions["action"]:
action_counter[a[0]] += 1
if max:
### if max, some action will always be chosen (which action is
### chosen depends on the network initialization
count = 0
for i in range(num_actions):
prob = action_counter[i] / float(sum(action_counter))
if abs(prob - 1.0) < 1e-1:
count = count + 1
self.assertEqual(count, 1)
else:
### the actions should be uniform
for i in range(num_actions):
prob = action_counter[i] / float(sum(action_counter))
self.assertAlmostEqual(prob, 1.0 / num_actions, places=1)
dims = 100
q_cnn = SimpleQ(model=TestModelCNN(
width=84, height=84, num_actions=num_actions))
q = SimpleQ(model=SimpleModelQ(
dims=[dims],
num_actions=num_actions,
perception_net=nn.Sequential(
nn.Linear(
dims, 32, bias=False),
nn.ReLU(),
nn.Linear(
32, 16, bias=False),
nn.ReLU())))
batch_size = 10
height, width = 84, 84
sensor = np.zeros([batch_size, dims]).astype("float32")
image = np.zeros([batch_size, 1, height, width]).astype("float32")
ct0 = ComputationTask("test", algorithm=q_cnn)
ct1 = ComputationTask("test", algorithm=q)
test(dict(image=image), ct0, max=False)
test(dict(sensor=sensor), ct1, max=True)
def test_predict(self):
"""
Test case for AC-learning and Q-learning predictions
"""
num_actions = 4
def test(input, ct, max):
action_counter = [0] * num_actions
total = 2000
for i in range(total):
actions, states = ct.predict(inputs=input)
assert not states, "states should be empty"
## actions["action"] is a batch of actions
for a in actions["action"]:
action_counter[a[0]] += 1
if max:
### if max, the first action will always be chosen
for i in range(num_actions):
prob = action_counter[i] / float(sum(action_counter))
self.assertAlmostEqual(prob,
1.0 if i == 0 else 0.0,
places=1)
else:
### the actions should be uniform
for i in range(num_actions):
prob = action_counter[i] / float(sum(action_counter))
self.assertAlmostEqual(prob, 1.0 / num_actions, places=1)
dims = 100
q_cnn = SimpleQ(
model=TestModelCNN(width=84, height=84, num_actions=num_actions))
q = SimpleQ(model=SimpleModelQ(
dims=dims,
num_actions=num_actions,
mlp=nn.Sequential(nn.Linear(dims, 32, bias=False), nn.ReLU(),
nn.Linear(32, 16, bias=False), nn.ReLU(),
nn.Linear(16, num_actions, bias=False))))
batch_size = 10
height, width = 84, 84
sensor = np.zeros([batch_size, dims]).astype("float32")
image = np.zeros([batch_size, 1, height, width]).astype("float32")
ct0 = ComputationTask(algorithm=q_cnn)
ct1 = ComputationTask(algorithm=q)
test(dict(image=image), ct0, max=False)
test(dict(sensor=sensor), ct1, max=True)
def test_ct_learning(self):
"""
Test training
"""
num_actions = 2
dims = 100
batch_size = 8
sensor = np.ones(
[batch_size, dims]).astype("float32") / dims # normalize
next_sensor = np.zeros([batch_size, dims]).astype("float32")
for on_policy in [True, False]:
if on_policy:
alg = SimpleAC(model=SimpleModelAC(
dims=dims,
num_actions=num_actions,
mlp=nn.Sequential(
nn.Linear(
dims, 64, bias=False),
nn.ReLU(),
nn.Linear(
64, 32, bias=False),
nn.ReLU())))
ct = ComputationTask(
"test", algorithm=alg, hyperparas=dict(lr=1e-1))
else:
alg = SimpleQ(
model=SimpleModelQ(
dims=dims,
num_actions=num_actions,
mlp=nn.Sequential(
nn.Linear(
dims, 64, bias=False),
nn.ReLU(),
nn.Linear(
64, 32, bias=False),
nn.ReLU(),
nn.Linear(
32, num_actions, bias=False))),
update_ref_interval=100)
ct = ComputationTask(
"test", algorithm=alg, hyperparas=dict(lr=1e-1))
for i in range(1000):
if on_policy:
outputs, _ = ct.predict(inputs=dict(sensor=sensor))
actions = outputs["action"]
else:
## randomly assemble a batch
actions = np.random.choice(
[0, 1], size=(batch_size, 1),
p=[0.5, 0.5]).astype("int")
rewards = (1.0 - actions).astype("float32")
cost = ct.learn(
inputs=dict(sensor=sensor),
next_inputs=dict(sensor=next_sensor),
next_alive=dict(alive=np.zeros(
(batch_size, 1)).astype("float32")),
actions=dict(action=actions),
rewards=dict(reward=rewards))
### the policy should bias towards the first action
outputs, _ = ct.predict(inputs=dict(sensor=sensor))
for a in outputs["action"]:
self.assertEqual(a[0], 0)
for _ in range(num_agents):
envs.append(GymEnv(game))
state_shape = envs[-1].observation_dims()[0]
num_actions = envs[-1].action_dims()[0]
# 2. Construct the network and specify the algorithm.
# Here we use a small MLP and apply the Q-learning algorithm
inner_size = 256
mlp = nn.Sequential(
nn.Linear(state_shape[0], inner_size),
nn.ReLU(),
nn.Linear(inner_size, inner_size),
nn.ReLU(), nn.Linear(inner_size, inner_size), nn.ReLU())
alg = SimpleQ(
model=SimpleModelQ(
dims=state_shape, num_actions=num_actions, perception_net=mlp),
exploration_end_steps=500000 / num_agents,
update_ref_interval=100)
# 3. Specify the settings for learning: the algorithm to use (SimpleAC
# in this case), data sampling strategy (ExpReplayHelper here) and other
# settings used by ComputationTask.
ct_settings = {
"RL": dict(
num_agents=num_agents,
algorithm=alg,
hyperparas=dict(lr=1e-4),
# sampling
agent_helper=ExpReplayHelper,
buffer_capacity=200000 / num_agents,
num_experiences=4, # num per agent
# 2. Construct the network and specify the algorithm.
# Here we use a small CNN as the perception net for the Actor-Critic algorithm
cnn = nn.Sequential(
nn.Conv2d(d, 32, kernel_size=8, stride=4),
nn.ReLU(),
nn.Conv2d(32, 64, kernel_size=4, stride=2),
nn.ReLU(),
nn.Conv2d(64, 64, kernel_size=3, stride=1),
nn.ReLU(),
Flatten(), # flatten the CNN cube to a vector
nn.Linear(7 * 7 * 64, 512),
nn.ReLU())
alg = SimpleQ(model=SimpleModelQ(dims=(d, h, w),
num_actions=num_actions,
perception_net=cnn),
gpu_id=0,
exploration_end_steps=500000 / num_agents,
update_ref_interval=100)
# 3. Specify the settings for learning: data sampling strategy
# (ExpReplayHelper here) and other settings used by
# ComputationTask.
ct_settings = {
"RL":
dict(
num_agents=num_agents,
algorithm=alg,
hyperparas=dict(lr=1e-4, grad_clip=5.0),
# sampling
def test_gym_games(self):
"""
Test games in OpenAI gym.
"""
games = ["MountainCar-v0", "CartPole-v0", "Pendulum-v0"]
final_rewards_thresholds = [
-1.5, ## drive to the right top in 150 steps (timeout is -2.0)
1.5, ## hold the pole for at least 150 steps
-3.0 ## can swing the stick to the top most of the times
]
on_policies = [False, True, False]
discrete_actions = [True, True, False]
for game, threshold, on_policy, discrete_action in \
zip(games, final_rewards_thresholds, on_policies, discrete_actions):
env = gym.make(game)
state_shape = env.observation_space.shape[0]
if discrete_action:
num_actions = env.action_space.n
else:
num_actions = env.action_space.shape[0]
hidden_size = 256
mlp = nn.Sequential(
nn.Linear(state_shape, hidden_size),
nn.ReLU(),
nn.Linear(hidden_size, hidden_size),
nn.ReLU(), nn.Linear(hidden_size, hidden_size), nn.ReLU())
q_model = SimpleModelQ(
dims=state_shape,
num_actions=num_actions,
mlp=nn.Sequential(mlp, nn.Linear(hidden_size, num_actions)))
if on_policy:
alg = SimpleSARSA(model=q_model, epsilon=0.1)
# alg = SuccessorRepresentationQ(
# ## much slower than SARSA because of more things to learn
# model=SimpleSRModel(
# dims=state_shape,
# hidden_size=hidden_size,
# num_actions=num_actions, ),
# exploration_end_steps=20000)
else:
if discrete_action:
alg = SimpleQ(
model=q_model,
exploration_end_steps=200000,
update_ref_interval=100)
else:
alg = OffPolicyAC(
model=GaussianPolicyModel(
dims=state_shape,
action_dims=num_actions,
mlp=mlp,
std=1.0),
epsilon=0.2)
glog.info("algorithm: " + alg.__class__.__name__)
ct = ComputationTask("RL", algorithm=alg, hyperparas=dict(lr=1e-4))
batch_size = 32
if not on_policy:
train_every_steps = batch_size / 4
buffer_size_limit = 200000
max_episode = 10000
average_episode_reward = []
past_exps = []
max_steps = env._max_episode_steps
for n in range(max_episode):
ob = env.reset()
episode_reward = 0
alive = 1
for t in range(max_steps):
inputs = dict(sensor=np.array([ob]).astype("float32"))
res, _ = ct.predict(inputs=inputs)
## when discrete_action is True, this is a scalar
## otherwise it's a floating vector
pred_action = res["action"][0]
## end before the env wrongly gives game_over=True for a timeout case
if t == max_steps - 1:
past_exps.append(
(inputs, res, dict(reward=[[0]]),
dict(alive=[[-1]]))) ## -1 denotes timeout
break
elif (not alive):
past_exps.append((inputs, res, dict(reward=[[0]]),
dict(alive=[[alive]])))
break
else:
next_ob, reward, next_is_over, _ = env.step(
pred_action[0] if discrete_action else pred_action)
reward /= 100
episode_reward += reward
past_exps.append((inputs, res, dict(reward=[[reward]]),
dict(alive=[[alive]])))
## only for off-policy training we use a circular buffer
if (not on_policy) and len(past_exps) > buffer_size_limit:
past_exps.pop(0)
## compute the learning condition
learn_cond = False
if on_policy:
learn_cond = (len(past_exps) >= batch_size)
else:
learn_cond = (
t % train_every_steps == train_every_steps - 1)
if learn_cond:
exps = sample(past_exps, batch_size)
sampled_inputs, next_sampled_inputs, sampled_actions, \
next_sampled_actions, reward, next_alive = unpack_exps(exps)
cost = ct.learn(
inputs=sampled_inputs,
next_inputs=next_sampled_inputs,
next_alive=next_alive,
actions=sampled_actions,
next_actions=next_sampled_actions,
rewards=reward)
## we clear the exp buffer for on-policy
if on_policy:
past_exps = []
ob = next_ob
### bool must be converted to int for correct computation
alive = 1 - int(next_is_over)
if n % 50 == 0:
glog.info("episode reward: %f" % episode_reward)
average_episode_reward.append(episode_reward)
if len(average_episode_reward) > 20:
average_episode_reward.pop(0)
### once hit the threshold, we don't bother running
if sum(average_episode_reward) / len(
average_episode_reward) > threshold:
glog.info(
"Test terminates early due to threshold satisfied!")
break
### compuare the average episode reward to reduce variance
self.assertGreater(
sum(average_episode_reward) / len(average_episode_reward),
threshold)
def test_gym_games(self):
"""
Test games in OpenAI gym.
"""
games = ["MountainCar-v0", "CartPole-v0"]
final_rewards_thresholds = [
-1.8, ## drive to the right top in 180 steps (timeout is -2.0)
1.5 ## hold the pole for at least 150 steps
]
for game, threshold in zip(games, final_rewards_thresholds):
for on_policy in [False, True]:
if on_policy and game != "CartPole-v0":
## SimpleAC has difficulty training mountain-car and acrobot
continue
env = gym.make(game)
state_shape = env.observation_space.shape[0]
num_actions = env.action_space.n
mlp = nn.Sequential(nn.Linear(state_shape, 128), nn.ReLU(),
nn.Linear(128, 128), nn.ReLU(),
nn.Linear(128, 128), nn.ReLU())
if on_policy:
alg = SimpleAC(model=SimpleModelAC(dims=state_shape,
num_actions=num_actions,
mlp=mlp),
hyperparas=dict(lr=1e-3))
else:
alg = SimpleQ(model=SimpleModelQ(
dims=state_shape,
num_actions=num_actions,
mlp=nn.Sequential(mlp, nn.Linear(128, num_actions))),
hyperparas=dict(lr=1e-4),
exploration_end_batches=25000,
update_ref_interval=100)
print "algorithm: " + alg.__class__.__name__
ct = ComputationTask(algorithm=alg)
batch_size = 16
if not on_policy:
train_every_steps = batch_size / 4
buffer_size_limit = 100000
max_episode = 5000
average_episode_reward = []
past_exps = []
max_steps = env._max_episode_steps
for n in range(max_episode):
ob = env.reset()
episode_reward = 0
for t in range(max_steps):
res, _ = ct.predict(inputs=dict(
sensor=np.array([ob]).astype("float32")))
pred_action = res["action"][0][0]
next_ob, reward, next_is_over, _ = env.step(
pred_action)
reward /= 100
episode_reward += reward
past_exps.append((ob, next_ob, [pred_action], [reward],
[not next_is_over]))
## only for off-policy training we use a circular buffer
if (not on_policy
) and len(past_exps) > buffer_size_limit:
past_exps.pop(0)
## compute the learning condition
learn_cond = False
if on_policy:
learn_cond = (len(past_exps) >= batch_size)
exps = past_exps ## directly use all exps in the buffer
else:
learn_cond = (
t % train_every_steps == train_every_steps - 1)
exps = sample(past_exps,
batch_size) ## sample some exps
if learn_cond:
sensor, next_sensor, action, reward, next_episode_end \
= unpack_exps(exps)
cost = ct.learn(
inputs=dict(sensor=sensor),
next_inputs=dict(next_sensor=next_sensor),
next_episode_end=dict(
next_episode_end=next_episode_end),
actions=dict(action=action),
rewards=dict(reward=reward))
## we clear the exp buffer for on-policy
if on_policy:
past_exps = []
ob = next_ob
## end before the Gym wrongly gives game_over=True for a timeout case
if t == max_steps - 2 or next_is_over:
break
if n % 50 == 0:
print("episode reward: %f" % episode_reward)
average_episode_reward.append(episode_reward)
if len(average_episode_reward) > 20:
average_episode_reward.pop(0)
### compuare the average episode reward to reduce variance
self.assertGreater(
sum(average_episode_reward) / len(average_episode_reward),
threshold)