def evaluate_MPC():
file_path = 'C://Users//REUBS_LEN//PycharmProjects//RocketLanding//control_and_ai//evaluation_scripts//mpc'
k, p = 100, 10
Q_weights = np.array([0.25, 4, 0.25, 0.25, 365, 131])
Q = np.eye(6)*Q_weights*k
R_weights = np.array([0.1, 0.1, 10])
R = np.eye(3)*R_weights*p
mpc_settings = {'time_horizon':30,
'control_horizon': 5,
'time_step': 1 / 30,
'time_span': 30,
'flight_time_span': 50,
'ground_altitude': 5.2,
'max_altitude': 26,
'x_target': 16.5,
'finite_differences_step_size': 50,
'Optimisation Type': 1,
'Q': Q,
'R': R}
simulation_settings['Observation Space Size'] = 6
env = RocketLander(simulation_settings)
controller = MPC_Controller(env, mpc_settings=mpc_settings)
testing_framework = Evaluation_Framework(simulation_settings, mpc_settings)
reward_results, final_state_history, action_history = testing_framework.execute_evaluation(env, controller,
INITIAL_STATES,
INITIAL_FORCES,
DISTURBANCES, IMPULSES)
return file_path, reward_results, final_state_history, action_history
def evaluate_pid():
file_path = 'C://Users//REUBS_LEN//PycharmProjects//RocketLanding//control_and_ai//evaluation_scripts//pid'
env = RocketLander(simulation_settings)
pid_controller = PID_Controller()
testing_framework = Evaluation_Framework(simulation_settings)
reward_results, final_state_history, action_history = testing_framework.execute_evaluation(env, pid_controller,
INITIAL_STATES,
INITIAL_FORCES,
DISTURBANCES, IMPULSES)
return file_path, reward_results, final_state_history, action_history
def evaluate_pid():
file_path = CWD + os.path.sep + 'evaluation_scripts' + os.path.sep + 'pid'
env = RocketLander(simulation_settings)
pid_controller = PID_Controller()
testing_framework = Evaluation_Framework(simulation_settings)
reward_results, final_state_history, action_history = testing_framework.execute_evaluation(env, pid_controller,
INITIAL_STATES,
INITIAL_FORCES,
DISTURBANCES, IMPULSES)
return file_path, reward_results, final_state_history, action_history
def evaluate_q_learning(model_settings):
load_path = CWD + os.path.sep + 'control_and_ai' + os.path.sep + 'rl_q_learning' + os.path.sep + model_settings['dir']
file_path = CWD + os.path.sep + 'evaluation_scripts' + os.path.sep + 'rl_q_learning' + os.path.sep + model_settings['dir']
env = RocketLander(simulation_settings)
controller = Q_Learning_Controller(load_path=load_path, low_discretization=False,
simulation_settings=simulation_settings)
testing_framework = Evaluation_Framework(simulation_settings)
reward_results, final_state_history, action_history = testing_framework.execute_evaluation(env, controller,
INITIAL_STATES,
INITIAL_FORCES,
DISTURBANCES, IMPULSES)
return file_path, reward_results, final_state_history, action_history
def evaluate_high_discretization_q_learning_function_approximation_longer_state():
file_path = 'C://Users//REUBS_LEN//PycharmProjects//RocketLanding//control_and_ai//evaluation_scripts//rl_q_learning//high_discretization_longer_state'
load_path = 'C://Users//REUBS_LEN//PycharmProjects//RocketLanding//control_and_ai//function_approximation_rl' \
'//rl_linear_function_approximation_high_function_discretization_5000_episodes_trained_at_once.p'
env = RocketLander(simulation_settings)
controller = Q_Learning_Controller_Longer_State(load_path=load_path, low_discretization=False,
simulation_settings=simulation_settings)
testing_framework = Evaluation_Framework(simulation_settings)
reward_results, final_state_history, action_history = testing_framework.execute_evaluation(env, controller,
INITIAL_STATES,
INITIAL_FORCES,
DISTURBANCES, IMPULSES)
return file_path, reward_results, final_state_history, action_history
def evaluate_unnormalized_longer_state_ddpg():
load_path = 'C://Users//REUBS_LEN//PycharmProjects//RocketLanding//control_and_ai//DDPG//model_2_longer_unnormalized_state'
file_path = 'C://Users//REUBS_LEN//PycharmProjects//RocketLanding//control_and_ai//evaluation_scripts//ddpg//model_2_unnormalized_longer_state'
normal_state_FLAG = False
untransformed_state_FLAG = False
normalized_FLAG = False
simulation_settings['Observation Space Size'] = 16
env = RocketLander(simulation_settings)
controller = Unnormalized_DDPG_Controller_Longer_State(env=env, load_path=load_path, normal_state_FLAG=normal_state_FLAG,
untransformed_state_FLAG=untransformed_state_FLAG,
normalized_FLAG=normalized_FLAG) # , simulation_settings=simulation_settings)
testing_framework = Evaluation_Framework(simulation_settings)
reward_results, final_state_history, action_history = testing_framework.execute_evaluation(env, controller,
INITIAL_STATES,
INITIAL_FORCES,
DISTURBANCES, IMPULSES)
return file_path, reward_results, final_state_history, action_history
def evaluate_ddpg(model_settings):
load_path = model_settings['dir']
file_path = CWD + os.path.sep + 'evaluation_scripts' + os.path.sep + 'ddpg' + os.path.sep + model_settings['dir']
short_state_FLAG = model_settings['short']
untransformed_state_FLAG = model_settings['untransformed']
normalized_FLAG = model_settings['normalized']
simulation_settings['Observation Space Size'] = model_settings['obs_size']
env = RocketLander(simulation_settings)
controller = DDPG_Controller(env=env, load_path=load_path, short_state_FLAG=short_state_FLAG,
untransformed_state_FLAG=untransformed_state_FLAG,
normalized_FLAG=normalized_FLAG)#, simulation_settings=simulation_settings)
testing_framework = Evaluation_Framework(simulation_settings)
reward_results, final_state_history, action_history = testing_framework.execute_evaluation(env, controller,
INITIAL_STATES,
INITIAL_FORCES,
DISTURBANCES, IMPULSES)
return file_path, reward_results, final_state_history, action_history
def evaluate_normalized_normal_state_ddpg():
# Model 1
# Fuel Cost = 0, Max Steps = 500, Episode Training = 2000, RANDOM FORCE = 20000, RANDOM X_FORCE = 0.2*RANDOM FORCE
load_path = 'C://Users//REUBS_LEN//PycharmProjects//RocketLanding//control_and_ai//DDPG//model_normal_state'
file_path = 'C://Users//REUBS_LEN//PycharmProjects//RocketLanding//control_and_ai//DDPG//model_normal_state'
normal_state_FLAG = True
untransformed_state_FLAG = False
normalized_FLAG = True
simulation_settings['Observation Space Size'] = 8
env = RocketLander(simulation_settings)
controller = Normalized_DDPG_Controller(env=env, load_path=load_path, normal_state_FLAG=normal_state_FLAG,
untransformed_state_FLAG=untransformed_state_FLAG,
normalized_FLAG=normalized_FLAG)#, simulation_settings=simulation_settings)
testing_framework = Evaluation_Framework(simulation_settings)
reward_results, final_state_history, action_history = testing_framework.execute_evaluation(env, controller,
INITIAL_STATES,
INITIAL_FORCES,
DISTURBANCES, IMPULSES)
return file_path, reward_results, final_state_history, action_history
eps = []
eps.append(OUPolicy(0, 0.2, 0.4))
eps.append(OUPolicy(0, 0.2, 0.4))
eps.append(OUPolicy(0, 0.2, 0.4))
simulation_settings = {'Side Engines': True,
'Clouds': True,
'Vectorized Nozzle': True,
'Graph': False,
'Render': False,
'Starting Y-Pos Constant': 1,
'Initial Force': 'random',
'Rows': 1,
'Columns': 2,
'Episodes': 500}
env = RocketLander(simulation_settings)
#env = wrappers.Monitor(env, '/tmp/contlunarlander', force=True, write_upon_reset=True)
FLAGS.retrain = False # Restore weights if False
FLAGS.test = False
FLAGS.num_episodes = 500
model_dir = 'C://Users//REUBS_LEN//PycharmProjects//RocketLanding//DDPG//model_2_longer_unnormalized_state'
with tf.device('/cpu:0'):
agent = DDPG(
action_bounds,
eps,
env.observation_space.shape[0],
actor_learning_rate=0.0001,
critic_learning_rate=0.001,
retrain=FLAGS.retrain,
log_dir=FLAGS.log_dir,