def experiment(n_episodes, default_policy=False, policy=None, render=False):
"""
Run a RL experiment that can be either training or testing
Args:
n_episodes: number of train/test episodes
default_policy: boolean to enable testing/training phase
policy: numpy tensor with a trained policy
render: enable OpenAI environment graphical rendering
Returns:
Dictionary with:
cumulative experiments outcomes
list of steps per episode
list of cumulative rewards
trained policy
"""
with tf.device('/gpu:0'):
res = [0, 0] # array of results accumulator: {[0]: Loss, [1]: Victory}
scores = [] # Cumulative rewards
steps = [] # Steps per episode
reward_list = RingBuffer(100)
env = gym.make('PongDeterministic-v4')
input_dim = env.observation_space.shape[0]
output_dim = env.action_space.n
if default_policy:
agent = DQNAgent(output_dim,
None,
use_ddqn=True,
default_policy=True,
model_filename=policy,
epsilon=0.05,
epsilon_lower_bound=0.05)
else:
layers = [
Conv2D(32, (8, 8),
strides=(4, 4),
activation='relu',
input_shape=(84, 84, 4),
kernel_initializer=VarianceScaling(scale=2.0)),
Conv2D(64, (4, 4),
strides=(2, 2),
activation='relu',
kernel_initializer=VarianceScaling(scale=2.0)),
Conv2D(64, (3, 3),
strides=(1, 1),
activation='relu',
kernel_initializer=VarianceScaling(scale=2.0)),
Flatten(),
Dense(512, activation='relu'),
Dense(output_dim)
]
agent = DQNAgent(output_dim,
layers,
use_ddqn=True,
memory_size=700000,
gamma=0.99,
learn_thresh=50000,
epsilon_lower_bound=0.02,
epsilon_decay_function=lambda e: e -
(0.98 / 950000),
update_rate=10000,
optimizer=Adam(0.00025))
gathered_frame = 0
for episode_number in tqdm(range(n_episodes), desc="Episode"):
frame = env.reset()
state = pre_processing(frame)
empty_state = np.zeros(state.shape, dtype="uint8")
cumulative_reward = 0
has_lost_life = True
t = 0
while True:
if has_lost_life:
next_action = 1 # [1, 4, 5][ran.randint(0, 2)]
stack = np.stack(
(empty_state, empty_state, empty_state, empty_state),
axis=2)
stack = np.reshape([stack], (1, 84, 84, 4))
for _ in range(ran.randint(1, 10)):
gathered_frame += 1
frame, reward, end, _ = env.step(next_action)
new_state = np.reshape(pre_processing(frame),
(1, 84, 84, 1))
new_stack = np.append(new_state,
stack[:, :, :, :3],
axis=3)
stack = new_stack
if (render):
env.render()
has_lost_life = False
next_action = agent.act(stack)
new_state, reward, end, _ = env.step(next_action)
if (render):
env.render()
time.sleep(0.02)
reward = np.clip(reward, -1., 1.)
if reward != 0:
has_lost_life = True
cumulative_reward += reward
new_state = np.reshape(pre_processing(new_state),
(1, 84, 84, 1))
new_stack = np.append(new_state, stack[:, :, :, :3], axis=3)
agent.memoise(
(stack, next_action, reward, new_state, has_lost_life))
stack = new_stack
gathered_frame += 1
if end:
reward_list.append(cumulative_reward)
if cumulative_reward > 0:
res[1] += 1
print("You Won!, steps:", t, "reward:",
reward_list.mean(), "frames:", gathered_frame)
else:
res[0] += 1
print("You Lost!, steps:", t, "reward:",
reward_list.mean(), "frames:", gathered_frame)
steps.append(t)
break
agent.learn()
t += 1
scores.append(cumulative_reward)
if episode_number >= 50 and episode_number % 10 == 0:
model_name = "partial_model_pong" + str(episode_number)
agent.save_model(model_name)
env.close()
return {
"results": np.array(res),
"steps": np.array(steps),
"scores": np.array(scores),
"agent": agent
}
def experiment(n_episodes, default_policy=False, policy=None, render=False):
"""
Run a RL experiment that can be either training or testing
Args:
n_episodes: number of train/test episodes
default_policy: boolean to enable testing/training phase
policy: numpy tensor with a trained policy
render: enable OpenAI environment graphical rendering
Returns:
Dictionary with:
cumulative experiments outcomes
list of steps per episode
list of cumulative rewards
trained policy
"""
res = [0, 0] # array of results accumulator: {[0]: Loss, [1]: Victory}
scores = [] # Cumulative rewards
steps = [] # Steps per episode
env = gym.make('MountainCar-v0')
env.seed(seed)
input_dim = env.observation_space.shape[0]
output_dim = env.action_space.n
layer1 = Dense(15, input_dim=input_dim, activation='relu')
layer2 = Dense(output_dim)
agent1 = DQNAgent(output_dim, [layer1, layer2],
use_ddqn=True,
learn_thresh=1000,
update_rate=300,
epsilon_decay_function=lambda e: e - 0.001,
epsilon_lower_bound=0.01,
optimizer=keras.optimizers.RMSprop(0.001),
tb_dir=None)
#agent2 = QLAgent([n_states, n_states, env.action_space.n], epsilon_decay_function=lambda e: e - 0.001, epsilon_lower_bound=0.01)
#agent3 = SARSAAgent([n_states, n_states, env.action_space.n], epsilon_decay_function=lambda e: e - 0.001, epsilon_lower_bound=0.01)
agent4 = DQNAgent(output_dim, [layer1, layer2],
use_ddqn=False,
learn_thresh=1000,
update_rate=300,
epsilon_decay_function=lambda e: e - 0.001,
epsilon_lower_bound=0.01,
optimizer=keras.optimizers.RMSprop(0.001),
tb_dir=None)
agents = [agent1, agent4]
agentE = EnsemblerAgent(env.action_space.n, agents,
EnsemblerType.TRUST_BASED)
evaluate = False
for i_episode in tqdm(range(n_episodes + 1), desc="Episode"):
state = env.reset()
# agent3.extract_policy()
discretized_state = obs_to_state(env, state, n_states)
cumulative_reward = 0
state = np.reshape(state, [1, 2])
if i_episode > 0 and i_episode % 100 == 0:
evaluate = True
if evaluate == False:
for t in range(env._max_episode_steps):
if (render):
env.render()
next_action = agentE.act(state, discretized_state)
new_state, reward, end, _ = env.step(next_action)
new_discretized_state = obs_to_state(env, new_state, n_states)
original_state = new_state
# Reward shaping
# r1 = reward + 0.1 * original_state[0]
# r2 = reward + 0.2 * np.sin(3 * original_state[0])
# r3 = reward + 0.7 * (original_state[1] * original_state[1])
r1 = reward + original_state[0]
r2 = reward + np.sin(3 * original_state[0])
r3 = reward + (original_state[1] * original_state[1])
r4 = abs(new_state[0] - (-0.5)) # r in [0, 1]
new_state = np.reshape(new_state, [1, 2])
agent1.memoise((state, next_action, r4, new_state, end))
#agent2.update_q((discretized_state[0], discretized_state[1]), (new_discretized_state[0], new_discretized_state[1]), next_action, reward)
#agent3.update_q((discretized_state[0], discretized_state[1]), (new_discretized_state[0], new_discretized_state[1]), next_action, reward)
agent4.memoise((state, next_action, r4, new_state, end))
if end:
if t == env._max_episode_steps - 1:
res[0] += 1
else:
res[1] += 1
print("ENTRATO!,", t, "steps", "reward: ",
cumulative_reward)
steps.append(t)
break
else:
state = new_state
discretized_state = new_discretized_state
cumulative_reward += reward
agent1.learn()
agent4.learn()
cumulative_reward += reward
scores.append(cumulative_reward)
else:
evaluate = False
eval_res = [
0, 0
] # array of results accumulator: {[0]: Loss, [1]: Victory}
eval_scores = [] # Cumulative rewards
eval_steps = [] # Steps per episode
for i_episode in range(100):
state = env.reset()
discretized_state = obs_to_state(env, state, n_states)
state = np.reshape(state, [1, 2])
cumulative_reward = 0
for t in range(env._max_episode_steps):
if (render):
env.render()
next_action = agentE.act(state, discretized_state)
new_state, reward, end, _ = env.step(next_action)
new_discretized_state = obs_to_state(
env, new_state, n_states)
original_state = new_state
new_state = np.reshape(new_state, [1, 2])
if end:
if t == env._max_episode_steps - 1:
eval_res[0] += 1
else:
eval_res[1] += 1
eval_steps.append(t)
break
else:
state = new_state
discretized_state = new_discretized_state
cumulative_reward += reward
cumulative_reward += reward
eval_scores.append(cumulative_reward)
testing_accuracy = accuracy(np.array(eval_res))
testing_mean_steps = np.array(eval_steps).mean()
testing_mean_score = np.array(eval_scores).mean()
print("\nTraining episodes:", len(steps), "Training mean score:",
np.array(steps).mean(), "Training mean steps",
np.array(scores).mean(), "\nAccuracy:", testing_accuracy,
"Test mean score:", testing_mean_score, "Test mean steps:",
testing_mean_steps)
env.close()
return {
"results": np.array(res),
"steps": np.array(steps),
"scores": np.array(scores)
}
def experiment(n_episodes, default_policy=False, policy=None, render=False):
"""
Run a RL experiment that can be either training or testing
Args:
n_episodes: number of train/test episodes
default_policy: boolean to enable testing/training phase
policy: numpy tensor with a trained policy
render: enable OpenAI environment graphical rendering
Returns:
Dictionary with:
cumulative experiments outcomes
list of steps per episode
list of cumulative rewards
trained policy
"""
res = [0, 0] # array of results accumulator: {[0]: Loss, [1]: Victory}
scores = [] # Cumulative rewards
steps = [] # steps per episode
env = gym.make('CartPole-v0')
env = env.unwrapped
env.seed(seed)
input_dim = env.observation_space.shape[0]
output_dim = env.action_space.n
layer1 = Dense(10, input_dim=input_dim, activation='relu')
layer2 = Dense(output_dim)
if default_policy:
agent = DQNAgent(output_dim,
None,
use_ddqn=True,
default_policy=True,
model_filename=policy,
epsilon=0,
epsilon_lower_bound=0,
learn_thresh=0)
else:
agent1 = DQNAgent(output_dim, [layer1, layer2],
use_ddqn=True,
learn_thresh=2000,
update_rate=100,
epsilon_decay_function=lambda e: e - 0.0001,
epsilon_lower_bound=0.1,
optimizer=keras.optimizers.RMSprop(0.001),
memory_size=2000,
tb_dir=None)
agent2 = DQNAgent(output_dim, [layer1, layer2],
use_ddqn=True,
learn_thresh=2000,
update_rate=100,
epsilon_decay_function=lambda e: e - 0.0001,
epsilon_lower_bound=0.1,
optimizer=keras.optimizers.RMSprop(0.001),
memory_size=2000,
tb_dir=None)
agent3 = DQNAgent(output_dim, [layer1, layer2],
use_ddqn=True,
learn_thresh=2000,
update_rate=100,
epsilon_decay_function=lambda e: e - 0.0001,
epsilon_lower_bound=0.1,
optimizer=keras.optimizers.RMSprop(0.001),
memory_size=2000,
tb_dir=None)
agent4 = DQNAgent(output_dim, [layer1, layer2],
use_ddqn=True,
learn_thresh=2000,
update_rate=100,
epsilon_decay_function=lambda e: e - 0.0001,
epsilon_lower_bound=0.1,
optimizer=keras.optimizers.RMSprop(0.001),
memory_size=2000,
tb_dir=None)
agent5 = DQNAgent(output_dim, [layer1, layer2],
use_ddqn=True,
learn_thresh=2000,
update_rate=100,
epsilon_decay_function=lambda e: e - 0.0001,
epsilon_lower_bound=0.1,
optimizer=keras.optimizers.RMSprop(0.001),
memory_size=2000,
tb_dir=None)
agent6 = DQNAgent(output_dim, [layer1, layer2],
use_ddqn=True,
learn_thresh=2000,
update_rate=100,
epsilon_decay_function=lambda e: e - 0.0001,
epsilon_lower_bound=0.1,
optimizer=keras.optimizers.RMSprop(0.001),
memory_size=2000,
tb_dir=None)
agent7 = DQNAgent(output_dim, [layer1, layer2],
use_ddqn=True,
learn_thresh=2000,
update_rate=100,
epsilon_decay_function=lambda e: e - 0.0001,
epsilon_lower_bound=0.1,
optimizer=keras.optimizers.RMSprop(0.001),
memory_size=2000,
tb_dir=None)
agent8 = DQNAgent(output_dim, [layer1, layer2],
use_ddqn=True,
learn_thresh=2000,
update_rate=100,
epsilon_decay_function=lambda e: e - 0.0001,
epsilon_lower_bound=0.1,
optimizer=keras.optimizers.RMSprop(0.001),
memory_size=2000,
tb_dir=None)
agent9 = DQNAgent(output_dim, [layer1, layer2],
use_ddqn=True,
learn_thresh=2000,
update_rate=100,
epsilon_decay_function=lambda e: e - 0.0001,
epsilon_lower_bound=0.1,
optimizer=keras.optimizers.RMSprop(0.001),
memory_size=2000,
tb_dir=None)
agent10 = DQNAgent(output_dim, [layer1, layer2],
use_ddqn=True,
learn_thresh=2000,
update_rate=100,
epsilon_decay_function=lambda e: e - 0.0001,
epsilon_lower_bound=0.1,
optimizer=keras.optimizers.RMSprop(0.001),
memory_size=2000,
tb_dir=None)
agents = [
agent1, agent2, agent3, agent4, agent5, agent6, agent7, agent8,
agent9, agent10
]
agentE = EnsemblerAgent(output_dim, agents, EnsemblerType.TRUST_BASED)
for i_ep in tqdm(range(n_episodes), desc="Episode"):
state = env.reset()
cumulative_reward = 0
state = np.reshape(state, [1, 4])
t = 0
while True:
if (render):
env.render()
time.sleep(0.1)
next_action = agentE.act(state)
new_state, reward, end, _ = env.step(next_action)
x, x_dot, theta, theta_dot = new_state
new_state = np.reshape(new_state, [1, 4])
# Reward shaping
r1 = (env.x_threshold - abs(x)) / env.x_threshold - 0.8
r2 = (env.theta_threshold_radians -
abs(theta)) / env.theta_threshold_radians - 0.5
r3 = -abs(theta_dot)
agent1.memoise((state, next_action, r2, new_state, end))
agent2.memoise((state, next_action, r3, new_state, end))
agent3.memoise((state, next_action, r2, new_state, end))
agent4.memoise((state, next_action, r3, new_state, end))
agent5.memoise((state, next_action, r2, new_state, end))
agent6.memoise((state, next_action, r3, new_state, end))
agent7.memoise((state, next_action, r2, new_state, end))
agent8.memoise((state, next_action, r3, new_state, end))
agent9.memoise((state, next_action, r2, new_state, end))
agent10.memoise((state, next_action, r3, new_state, end))
if end or t > 199:
if t < 195:
res[0] += 1
else:
res[1] += 1
print("ENTRATO!,", t, "steps", "reward: ",
cumulative_reward)
steps.append(t)
if i_ep % 100 == 0:
if evaluate(env, agentE):
cumulative_reward += reward
scores.append(cumulative_reward)
env.close()
return {
"results": np.array(res),
"steps": np.array(steps),
"scores": np.array(scores)
}
break
else:
state = new_state
cumulative_reward += reward
for agent in agentE.agents:
agent.learn()
t += 1
cumulative_reward += reward
scores.append(cumulative_reward)
env.close()
return {
"results": np.array(res),
"steps": np.array(steps),
"scores": np.array(scores)
}
def experiment(n_episodes,
default_policy=False,
policy=None,
render=False,
agent_config=None):
"""
Run a RL experiment that can be either training or testing
Args:
n_episodes: number of train/test episodes
default_policy: boolean to enable testing/training phase
policy: numpy tensor with a trained policy
render: enable OpenAI environment graphical rendering
agent_config: DQNAgent object
Returns:
Dictionary with:
cumulative experiments outcomes
list of steps per episode
list of cumulative rewards
trained policy
"""
res = [0, 0] # array of results accumulator: {[0]: Loss, [1]: Victory}
scores = [] # Cumulative rewards
steps = [] # Steps per episode
env = gym.make('MountainCar-v0')
env.seed(seed)
input_dim = env.observation_space.shape[0]
output_dim = env.action_space.n
if agent_config is None:
if default_policy:
agent = DQNAgent(output_dim,
None,
use_ddqn=True,
default_policy=True,
model_filename=policy,
epsilon=0,
epsilon_lower_bound=0,
learn_thresh=0)
else:
layer1 = Dense(15, input_dim=input_dim, activation='relu')
layer2 = Dense(output_dim)
agent = DQNAgent(output_dim, [layer1, layer2],
use_ddqn=True,
learn_thresh=1000,
update_rate=300,
epsilon_decay_function=lambda e: e * 0.95,
epsilon_lower_bound=0.01,
optimizer=keras.optimizers.RMSprop(0.001))
else:
agent = agent_config
for i_episode in tqdm(range(n_episodes), desc="Episode"):
state = env.reset()
cumulative_reward = 0
# Model validation for early stopping
if i_episode > 0 and (i_episode % 100) == 0 and not default_policy:
agent.save_model("tmp_model")
evaluation_result = experiment(500,
default_policy=True,
policy="tmp_model")
acc = accuracy(evaluation_result["results"])
if acc == 100:
break
else:
print("Accuracy:", acc, "Episode:", i_episode)
state = np.reshape(state, [1, 2])
for t in range(env._max_episode_steps):
if (render):
env.render()
next_action = agent.act(state)
new_state, reward, end, _ = env.step(next_action)
reward = abs(new_state[0] - (-0.5)) # r in [0, 1] (reward shaping)
new_state = np.reshape(new_state, [1, 2])
agent.memoise((state, next_action, reward, new_state, end))
if end:
if t == env._max_episode_steps - 1:
res[0] += 1
else:
res[1] += 1
# print("ENTRATO!,", t, "steps")
steps.append(t)
break
else:
state = new_state
cumulative_reward += reward
agent.learn()
cumulative_reward += reward
scores.append(cumulative_reward)
env.close()
return {
"results": np.array(res),
"steps": np.array(steps),
"scores": np.array(scores),
"agent": agent
}