VRL_CodeReview/pendulum_change_l.py at master · rowangithub/VRL_CodeReview · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
# -*- coding: utf-8 -*-
# -------------------------------
# Author: Zikang Xiong
# Email: zikangxiong@gmail.com
# Date:   2018-10-23 17:04:25
# Last Modified by:   Zikang Xiong
# Last Modified time: 2019-02-22 16:44:32
# -------------------------------
from main import *
import numpy as np
from DDPG import *

from shield import Shield
from Environment import Environment
import argparse

def pendulum(learning_eposides, critic_structure, actor_structure, train_dir, learning_method, number_of_rollouts, simulation_steps,\
            nn_test=False, retrain_shield=False, shield_test=False, test_episodes=100):

  m = 1.
  l = 1.2
  g = 10.

  #Dynamics that are continuous
  A = np.matrix([
    [ 0., 1.],
    [g/l, 0.]
    ])
  B = np.matrix([
    [          0.],
    [1./(m*l**2.)]
    ])


  #intial state space
  s_min = np.array([[-0.35],[-0.35]])
  s_max = np.array([[ 0.35],[ 0.35]])

  #reward function
  Q = np.matrix([[1., 0.],[0., 1.]])
  R = np.matrix([[.005]])

  #safety constraint
  x_min = np.array([[-0.5],[-0.5]])
  x_max = np.array([[ 0.5],[ 0.5]])
  u_min = np.array([[-15.]])
  u_max = np.array([[ 15.]])

  env = Environment(A, B, u_min, u_max, s_min, s_max, x_min, x_max, Q, R, continuous=True)

  args = { 'actor_lr': 0.0001,
           'critic_lr': 0.001,
           'actor_structure': actor_structure,
           'critic_structure': critic_structure,
           'buffer_size': 1000000,
           'gamma': 0.99,
           'max_episode_len': 1,
           'max_episodes': learning_eposides,
           'minibatch_size': 64,
           'random_seed': 6553,
           'tau': 0.005,
           'model_path': train_dir+"model.chkp",
           'enable_test': nn_test,
           'test_episodes': test_episodes,
           'test_episodes_len': 3000}

  actor = DDPG(env, args)

  #################### Shield #################
  model_path = os.path.split(args['model_path'])[0]+'/'
  linear_func_model_name = 'K.model'
  model_path = model_path+linear_func_model_name+'.npy'

  def rewardf(x, Q, u, R):
    return env.reward(x, u)

  shield = Shield(env, actor, model_path, force_learning=retrain_shield, debug=False)
  shield.train_shield(learning_method, number_of_rollouts, simulation_steps, rewardf=rewardf, eq_err=1e-2, explore_mag = 0.3, step_size = 0.3)
  if shield_test:
    shield.test_shield(test_episodes, 3000, mode="single")

  actor.sess.close()

if __name__ == "__main__":
  parser = argparse.ArgumentParser(description='Running Options')
  parser.add_argument('--nn_test', action="store_true", dest="nn_test")
  parser.add_argument('--retrain_shield', action="store_true", dest="retrain_shield")
  parser.add_argument('--shield_test', action="store_true", dest="shield_test")
  parser.add_argument('--test_episodes', action="store", dest="test_episodes", type=int)
  parser_res = parser.parse_args()
  nn_test = parser_res.nn_test
  retrain_shield = parser_res.retrain_shield
  shield_test = parser_res.shield_test
  test_episodes = parser_res.test_episodes if parser_res.test_episodes is not None else 100

  pendulum(0, [1200,900], [1000,900,800], "ddpg_chkp/perfect_model/pendulum/change_l/", "random_search", 100, 2000, nn_test=nn_test, retrain_shield=retrain_shield, shield_test=shield_test, test_episodes=test_episodes)