def __init__(self, args, env, sess):
self.firstIter = 1
self.count = 1
self.args = args
self.sess = sess
self.env = env
self.torque_bound = 15.
self.max_speed = 60.
#Set up observation space and action space
self.observation_space = env.observation_space
self.action_space = env.action_space
print('Observation space', self.observation_space)
print('Action space', self.action_space)
#Determine dimensions of observation & action space
self.observation_size = self.env.observation_space.shape[0]
self.action_size = self.action_space.shape[0]
# Build neural network model for observations/actions
self.build_model()
# Build barrier function model
cbf.build_barrier(self)
# Build GP model of dynamics
dynamics_gp.build_GP_model(self)
def main(args, reward_result):
with tf.Session() as sess:
env = gym.make(args['env'])
np.random.seed(int(args['random_seed']))
tf.set_random_seed(int(args['random_seed']))
env.seed(int(args['random_seed']))
# Set environment parameters for pendulum
env.unwrapped.max_torque = 15.
env.unwrapped.max_speed = 60.
env.unwrapped.action_space = spaces.Box(low=-env.unwrapped.max_torque,
high=env.unwrapped.max_torque,
shape=(1, ))
high = np.array([1., 1., env.unwrapped.max_speed])
env.unwrapped.observation_space = spaces.Box(low=-high, high=high)
state_dim = env.observation_space.shape[0]
action_dim = env.action_space.shape[0]
action_bound = env.action_space.high
# Ensure action bound is symmetric
assert (env.action_space.high == -env.action_space.low)
actor = ActorNetwork(sess, state_dim, action_dim, action_bound,
float(args['actor_lr']), float(args['tau']),
int(args['minibatch_size']))
critic = CriticNetwork(sess, state_dim, action_dim,
float(args['critic_lr']), float(args['tau']),
float(args['gamma']),
actor.get_num_trainable_vars())
actor_noise = OrnsteinUhlenbeckActionNoise(mu=np.zeros(action_dim))
agent = LEARNER(env)
cbf.build_barrier(agent)
dynamics_gp.build_GP_model(agent)
agent.bar_comp = BARRIER(sess, 3, 1)
[summary_ops, summary_vars,
paths] = train(sess, env, args, actor, critic, actor_noise,
reward_result, agent)
return [summary_ops, summary_vars, paths]
def __init__(self, args, env, sess):
self.args = args
self.sess = sess
self.env = env
self.firstIter = 1
self.torque_bound = 100
#Determine dimensions of observation & action space
self.observation_size = 15
self.action_size = 1
# Build neural network model for observations/actions
self.build_model()
# Build barrier function model
cbf.build_barrier(self)
# Build GP model
dynamics_gp.build_GP_model(self)
def __init__(self, env, sess):
self.firstIter = 1
self.count = 1
self.env = env
self.torque_bound = 100.
'''
#Set up observation space and action space
self.observation_space = env.observation_space
self.action_space = env.action_space
print('Observation space', self.observation_space)
print('Action space', self.action_space)
'''
#Determine dimensions of observation & action space
self.observation_size = 15
self.action_size = 1
# Build barrier function model
cbf.build_barrier(self)
# Build GP model of dynamics
dynamics_gp.build_GP_model(self)
self.bar_comp = BARRIER(sess, 15, 1)
def main(args, reward_result, log_path):
with tf.Session() as sess:
env = gym.make(args['env'])
np.random.seed(int(args['random_seed']))
tf.set_random_seed(int(args['random_seed']))
env.seed(int(args['random_seed']))
# Create the log files
if not os.path.isdir(log_path):
os.mkdir(log_path)
log_save_name = log_path + '/episode_performance.csv'
f = open(log_save_name, "w+")
f.write(
"episode number, steps in evaluation, accumulated reward, done \n")
f.close()
log_save_name_cbf = log_path + '/episode_cbf_performance.csv'
f = open(log_save_name_cbf, "w+")
f.write(
"episode number, steps in evaluation, accumulated reward, done \n")
f.close()
# Set environment parameters for pendulum
env.unwrapped.max_torque = 15.
env.unwrapped.max_speed = 60.
env.unwrapped.action_space = spaces.Box(low=-env.unwrapped.max_torque,
high=env.unwrapped.max_torque,
shape=(1, ))
high = np.array([1., 1., env.unwrapped.max_speed])
env.unwrapped.observation_space = spaces.Box(low=-high, high=high)
state_dim = env.observation_space.shape[0]
action_dim = env.action_space.shape[0]
action_bound = env.action_space.high
# Ensure action bound is symmetric
assert (env.action_space.high == -env.action_space.low)
actor = ActorNetwork(sess, state_dim, action_dim, action_bound,
float(args['actor_lr']), float(args['tau']),
int(args['minibatch_size']))
critic = CriticNetwork(sess, state_dim, action_dim,
float(args['critic_lr']), float(args['tau']),
float(args['gamma']),
actor.get_num_trainable_vars())
actor_noise = OrnsteinUhlenbeckActionNoise(mu=np.zeros(action_dim))
agent = LEARNER(env)
cbf.build_barrier(agent)
dynamics_gp.build_GP_model(agent)
agent.bar_comp = BARRIER(sess, 3, 1)
[summary_ops, summary_vars,
paths] = train(sess, env, args, actor, critic, actor_noise,
reward_result, agent, log_save_name, log_save_name_cbf)
# Evaluate the final model 100 times to get a better idea of the final model's performance
f = open(args['log_path'] + '/final_eval.csv', "w+")
f.write("reward, steps, done, safe\n")
episode_length = int(args['max_episode_len'])
for k in range(100):
steps, reward, done, safe = evaluate(env, actor, episode_length)
f.write(
str(reward) + ', ' + str(steps) + ', ' + str(done) + ', ' +
str(safe) + '\n')
f.close()
# Evaluate the final model 100 times to get a better idea of the final model's performance
f = open(args['log_path'] + '/final_cbf_eval.csv', "w+")
f.write("reward, steps, done, safe\n")
episode_length = int(args['max_episode_len'])
for k in range(100):
steps, reward, done, safe = evaluate_with_cbf(
env, actor, agent, episode_length)
f.write(
str(reward) + ', ' + str(steps) + ', ' + str(done) + ', ' +
str(safe) + '\n')
f.close()
return [summary_ops, summary_vars, paths]