def __init__(self):
dat = loadmat('car_data_formatted_arc')
x = np.copy(np.squeeze(dat['car_dat']))
aaa = np.arange(len(x))
random.shuffle(aaa)
self.data = x[aaa]
y = np.copy(np.squeeze(dat['car_dat']))
self.data_orig = y[aaa]
self.count = 0
self.episode = -1
self.L = 100
self.numCars = 5
self.dt = 0.1
self.collision_flag = 0
self.state = np.copy(np.squeeze(self.data[0][0]))
self.bot_state = np.copy(np.squeeze(self.data[0][0][9:12]))
self.prior = BasePrior()
self.action_space = spaces.Box(low=-7.0, high = 3.0, shape = (1,))
high = np.array([
np.finfo(np.float32).max,
np.finfo(np.float32).max,
np.finfo(np.float32).max,
np.finfo(np.float32).max,
np.finfo(np.float32).max,
np.finfo(np.float32).max])
self.observation_space = spaces.Box(-high, high)
def __init__(self, args, sess):
self.args = args
self.sess = sess
[A, B] = get_linear_dynamics()
self.prior = BasePrior(A, B)
self.env = gym.make(self.args.env_name)
self.args.max_path_length = self.env.spec.timestep_limit
self.agent = TRPO(self.args, self.env, self.sess, self.prior)
def __init__(self):
dat = loadmat('car_data_formatted_arc')
x = np.copy(np.squeeze(dat['car_dat']))
aaa = np.arange(len(x))
random.shuffle(aaa)
self.data = x[aaa]
y = np.copy(np.squeeze(dat['car_dat']))
self.data_orig = y[aaa]
self.count = 0
self.episode = -1
self.L = 100
self.numCars = 5
self.dt = 0.1
self.collision_flag = 0
self.state = np.copy(np.squeeze(self.data[0][0]))
self.bot_state = np.copy(np.squeeze(self.data[0][0][9:12]))
self.prior = BasePrior()
def train(self, replay_buffer, minibatch_size):
# Get dynamics and initialize prior controller
prior = BasePrior()
#self.sess.as_default()
# Needed to enable BatchNorm
#tflearn.is_training(True)
#Sample a batch from the replay buffer
s_batch, a_batch, r_batch, t_batch, s2_batch = replay_buffer.sample_batch(
minibatch_size)
# Calculate targets
target_q = self.critic.predict_target(
s2_batch, self.actor.predict_target(s2_batch))
y_i = []
for k in range(minibatch_size):
if t_batch[k]:
y_i.append(r_batch[k])
else:
y_i.append(r_batch[k] + self.critic.gamma * target_q[k])
# Update the critic given the targets
predicted_q_value, _ = self.critic.train(
s_batch, a_batch, np.reshape(y_i, (minibatch_size, 1)))
ep_ave_max_q += np.amax(predicted_q_value)
# Update the actor policy using the sampled gradient
a_outs = self.actor.predict(s_batch)
grads = self.critic.action_gradients(s_batch, a_outs)
self.actor.train(s_batch, grads[0])
# Update target networks
self.actor.update_target_network()
self.critic.update_target_network()
def train(sess, env, args, actor, critic, actor_noise, reward_result):
# Set up summary Ops
summary_ops, summary_vars = build_summaries()
sess.run(tf.global_variables_initializer())
# Get dynamics and initialize prior controller
prior = BasePrior()
# Initialize target network weights
actor.update_target_network()
critic.update_target_network()
# Initialize replay memory
replay_buffer = ReplayBuffer(int(args['buffer_size']),
int(args['random_seed']))
# Needed to enable BatchNorm.
tflearn.is_training(True)
paths = list()
lambda_store = np.zeros((int(args['max_episode_len']), 1))
for i in range(int(args['max_episodes'])):
s = env.reset_inc()
ep_reward = 0.
ep_ave_max_q = 0
obs, action, act_prior, rewards, obs_ref, prior_ref, collisions = [], [], [], [], [], [], []
#Get reward using baseline controller
s0 = np.copy(s)
ep_reward_opt = 0.
for kk in range(int(args['max_episode_len'])):
a = env.getPrior()
prior_ref.append(np.array([a]))
s0, r, stop_c, act = env.step(a)
ep_reward_opt += r
obs_ref.append(s0)
if (stop_c):
break
# Get reward using regRL algorithm
s = env.reset()
for j in range(int(args['max_episode_len'])):
# Set control prior regularization weight
lambda_mix = 15.
lambda_store[j] = lambda_mix
# Get control prior
a_prior = env.getPrior()
# Rl control with exploration noise
ab = actor.predict(np.reshape(s, (1, actor.s_dim))) + actor_noise()
# Mix the actions (RL controller + control prior)
act = ab[0] / (1 + lambda_mix) + (lambda_mix /
(1 + lambda_mix)) * a_prior
# Take action and observe next state/reward
s2, r, terminal, act = env.step(act)
collisions.append(env.collision_flag)
act = np.array(act, ndmin=1)
# Add info from time step to the replay buffer
replay_buffer.add(np.reshape(s, (actor.s_dim, )),
np.reshape(ab, (actor.a_dim, )), r, terminal,
np.reshape(s2, (actor.s_dim, )),
np.reshape(a_prior, (actor.a_dim, )))
# Keep adding experience to the memory until
# there are at least minibatch size samples
if replay_buffer.size() > int(args['minibatch_size']):
#Sample a batch from the replay buffer
s_batch, a_batch_0, r_batch, t_batch, s2_batch, a_prior_batch = \
replay_buffer.sample_batch(int(args['minibatch_size']))
a_batch = a_batch_0
# Calculate targets
target_q = critic.predict_target(
s2_batch, actor.predict_target(s2_batch))
y_i = []
for k in range(int(args['minibatch_size'])):
if t_batch[k]:
y_i.append(r_batch[k])
else:
y_i.append(r_batch[k] + critic.gamma * target_q[k])
# Update the critic given the targets
predicted_q_value, _ = critic.train(
s_batch, a_batch,
np.reshape(y_i, (int(args['minibatch_size']), 1)))
ep_ave_max_q += np.amax(predicted_q_value)
# Update the actor policy using the sampled gradient
a_outs = actor.predict(s_batch)
grads = critic.action_gradients(s_batch, a_outs)
actor.train(s_batch, grads[0])
# Update target networks
actor.update_target_network()
critic.update_target_network()
s = s2
ep_reward += r
obs.append(s)
rewards.append(r)
action.append(act)
act_prior.append(np.array([a_prior]))
# Collect results at end of episode
if terminal:
print('| Reward: {:d} | Episode: {:d} | Qmax: {:.4f}'.format(int(ep_reward-ep_reward_opt), \
i, (ep_ave_max_q / float(j))))
reward_result[0, i] = ep_reward
reward_result[1, i] = ep_reward_opt
reward_result[2, i] = np.mean(lambda_store)
reward_result[3, i] = max(collisions)
path = {
"Observation": np.concatenate(obs).reshape((-1, 6)),
"Observation_ref": np.concatenate(obs_ref).reshape(
(-1, 6)),
"Action": np.concatenate(action),
"Action_Prior": np.concatenate(act_prior),
"Action_Prior_Ref": np.concatenate(prior_ref),
"Reward": np.asarray(rewards)
}
paths.append(path)
break
return [summary_ops, summary_vars, paths]
def train(sess, env, args, actor, critic, actor_noise, reward_result):
# Set up summary Ops
summary_ops, summary_vars = build_summaries()
sess.run(tf.global_variables_initializer())
# Get dynamics and initialize prior controller
[A, B] = get_linear_dynamics()
prior = BasePrior(A, B)
# Initialize target network weights
actor.update_target_network()
critic.update_target_network()
# Initialize replay memory
replay_buffer = ReplayBuffer(int(args['buffer_size']),
int(args['random_seed']))
paths = list()
for i in range(int(args['max_episodes'])):
s = env.reset()
ep_reward = 0.
ep_ave_max_q = 0
obs, action, rewards = [], [], []
#Get optimal reward using optimal control
s0 = np.copy(s)
ep_reward_opt = 0.
for kk in range(int(args['max_episode_len'])):
a_prior = prior.getControl_h(s0)
a = a_prior
s0, r, stop_c, _ = env.step(a)
ep_reward_opt += r
if (stop_c):
break
# Get reward using regRL algorithm
env.reset()
s = env.unwrapped.reset(s)
for j in range(int(args['max_episode_len'])):
# Set control prior regularization weight
lambda_mix = 5.
# Prior control
a_prior = prior.getControl_h(s)
# Rl control with exploration noise
a = actor.predict(np.reshape(s, (1, actor.s_dim))) + actor_noise()
#a = actor.predict(np.reshape(s, (1, actor.s_dim))) + (1. / (1. + i))
# Mix the actions (RL controller + control prior)
act = a[0] / (1 + lambda_mix) + (lambda_mix /
(1 + lambda_mix)) * a_prior
# Take action and observe next state/reward
s2, r, terminal, info = env.step(act)
# Add info from time step to the replay buffer
replay_buffer.add(
np.reshape(s, (actor.s_dim, )), np.reshape(a, (actor.a_dim, )),
r, terminal, np.reshape(s2, (actor.s_dim, )),
np.reshape((lambda_mix / (1 + lambda_mix)) * a_prior,
(actor.a_dim, )))
# Keep adding experience to the memory until
# there are at least minibatch size samples
if replay_buffer.size() > int(args['minibatch_size']):
#Sample a batch from the replay buffer
s_batch, a_batch_0, r_batch, t_batch, s2_batch, a_prior_batch = \
replay_buffer.sample_batch(int(args['minibatch_size']))
a_batch = a_batch_0
# Calculate targets
target_q = critic.predict_target(
s2_batch, actor.predict_target(s2_batch))
y_i = []
for k in range(int(args['minibatch_size'])):
if t_batch[k]:
y_i.append(r_batch[k])
else:
y_i.append(r_batch[k] + critic.gamma * target_q[k])
# Update the critic given the targets
predicted_q_value, _ = critic.train(
s_batch, a_batch,
np.reshape(y_i, (int(args['minibatch_size']), 1)))
ep_ave_max_q += np.amax(predicted_q_value)
# Update the actor policy using the sampled gradient
a_outs = actor.predict(s_batch)
grads = critic.action_gradients(s_batch, a_outs)
actor.train(s_batch, grads[0])
# Update target networks
actor.update_target_network()
critic.update_target_network()
# Calculate TD-Error for each state
base_q = critic.predict_target(s_batch,
actor.predict_target(s_batch))
target_q = critic.predict_target(
s2_batch, actor.predict_target(s2_batch))
s = s2
ep_reward += r
obs.append(s)
rewards.append(r)
action.append(a[0])
# Collect results at end of episode
if terminal:
for ii in range(len(obs)):
obs[ii] = obs[ii].reshape((4, 1))
print('| Reward: {:d} | Episode: {:d} | Qmax: {:.4f}'.format(int(ep_reward - ep_reward_opt), \
i, (ep_ave_max_q / float(j))))
reward_result[0, i] = ep_reward
reward_result[1, i] = ep_reward_opt
path = {
"Observation": np.concatenate(obs).reshape((-1, 4)),
"Action": np.concatenate(action),
"Reward": np.asarray(rewards)
}
paths.append(path)
print(ep_reward)
break
return [summary_ops, summary_vars, paths]
TIMESTAMP)
#env = gym.make(ENVIRONMENT)
env = allCars()
#env = wrappers.Monitor(env, os.path.join(SUMMARY_DIR, ENVIRONMENT), video_callable=None)
ppo = PPO(env, SUMMARY_DIR, gpu=True)
if MODEL_RESTORE_PATH is not None:
ppo.restore_model(MODEL_RESTORE_PATH)
t, terminal = 0, False
buffer_s, buffer_a, buffer_r, buffer_v, buffer_terminal = [], [], [], [], []
rolling_r = RunningStats()
# Get prior and set tuning parameters for adaptive regularization weight
prior = BasePrior()
lambda_store = np.zeros(BATCH + 1)
lambda_all = np.zeros(EP_MAX + 1)
lambda_max = 8
factor = 0.2
reward_total, reward_diff = [], []
for episode in range(EP_MAX + 1):
# Baseline reward using only control prior
sp = env.reset_inc()
reward_prior = 0.
while True:
a_prior = env.getPrior()
sp, reward_p, done_p, _ = env.step(a_prior)
TIMESTAMP = datetime.now().strftime("%Y%m%d-%H%M%S")
SUMMARY_DIR = os.path.join(OUTPUT_RESULTS_DIR, "PPO", ENVIRONMENT, TIMESTAMP)
env = gym.make(ENVIRONMENT)
ppo = PPO(env, SUMMARY_DIR, gpu=True)
if MODEL_RESTORE_PATH is not None:
ppo.restore_model(MODEL_RESTORE_PATH)
t, terminal = 0, False
buffer_s, buffer_a, buffer_r, buffer_v, buffer_terminal = [], [], [], [], []
rolling_r = RunningStats()
# Initialize control prior
[A,B] = get_linear_dynamics()
prior = BasePrior(A,B)
# Set fixed regularization weight
# lambda_mix = 4.
reward_total, reward_diff, reward_lqr_prior, reward_h_prior = [], [], [], []
for episode in range(EP_MAX + 1):
# Baseline reward using only control prior
s0 = env.reset()
sp = np.copy(s0)
reward_prior = 0.
while True:
a_prior = prior.getControl_h(sp)
a_prior = np.squeeze(np.asarray(a_prior))
sp, reward_p, done_p, _ = env.step(a_prior)