-
Notifications
You must be signed in to change notification settings - Fork 0
/
Environment_old.py
208 lines (157 loc) · 6.74 KB
/
Environment_old.py
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
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
import gym
from gym.spaces import Box
from Dynamics import state_to_coords, get_next_state, normalize_angle
import numpy as np
from matplotlib import animation
import matplotlib.pyplot as plt
# normalized angle
class ObservationSpaceCartPole():
def __init__(self):
self.shape = (6,)
class ActionSpaceCartPole():
def __init__(self):
self.shape = (1,)
self.bounds = (-20, 20)
class DoublePendulumEnv(gym.Env):
def __init__(self, init_state, dt=0.02, max_initial_angle = 3 * 2 * np.pi / 360, discretized_actions = False):
if discretized_actions:
self.action_space = np.array([-10, 10, -5, 5, -1, 1, 0, 2, -2, 4, -4, 3, -3])
else:
self.action_space = Box(np.array([-10]), np.array([10]), dtype = np.float64)
self.observation_space = ObservationSpaceCartPole()
self.state = init_state
self.init_state = init_state
self.dt = dt
print('Environment initialized')
self.init_coords = state_to_coords(init_state)
self.max_initial_angle = max_initial_angle
self.state_history = [self.init_state]
self.action_history = []
def _take_action(self, action):
self.state = get_next_state(self.state, action, self.dt)
def _reward_function(self):
"""
# Reward system 1
Check whether 1 and 2 cart pole are in angle range between 80 and 100 degrees
agent will agent a reward in range [0, 1]
else:
If angle of pole 1 and 2 are lower than 80 and higher than 100 degrees, therefore, it makes sense
to terminate the environment and reset/restart.
agent will get a reward = -100
# Reward system 2
If cart is out of a given range of x = [-2, 2] then terminates the environment and
penalize the system heavily of a penalty = -100 and system is done here.
# Reward system 3
Penalize the system if spinning to fast
"""
done = False
state = self.state
reward = 0
# degree reward
normalized_angle_1 = np.degrees((state[1]))
normalized_angle_2 = np.degrees((state[2]))
#
if normalized_angle_1 > 80 and normalized_angle_1 < 95:
reward = 1 - (90 - normalized_angle_1) * 0.01
if normalized_angle_2 > 85 and normalized_angle_2 < 95:
reward += reward + 1 - (90 - normalized_angle_2) * 0.01
reward *= 4
else:
reward = -100
done = True
# another degree reward system
# cost = 2*(normalize_angle(state[1])/2 - np.pi/2) + \
# 2*(normalize_angle(state[2])/2 - np.pi/2)
#
# reward = -np.abs(cost)
# another degree_reward system
# deg_reward = ((np.sin(state[1]))*10 + (np.sin(state[2]))*10)/2
# #if np.sin(state[1]
# reward += deg_reward
# print(state[1])
# distance penalty
if state[0] > 1 or state[0] < -1:
reward -= 100
done = True
# distance2 rew
# state_coords = state_to_coords(state)
# # dist_pen = (state_coords[0][1] - state_coords[0][0])**2 + (state_coords[0][2] - state_coords[0][0])**2
# dist_rew = -( state_coords[1][1] - self.init_coords[1][1]) - ( state_coords[1][2] - self.init_coords[1][2])*2
# reward -= dist_rew
vel_reward = np.abs(state[4]*10) #minus points - we dont want it to spin super fast
reward -= vel_reward
# print(state[4]*10)
return reward, done
def reward_function3(self):
goal = np.array([0, 0])
coords = np.array(state_to_coords(self.state)[:, 0])
reward = max(np.linalg.norm(coords - goal), 0.0001)
reward = 1 / reward
done = False
return reward, done
def _reward_function2(self):
final_node_coords = np.array(state_to_coords(self.state)[:, -1])
goal_coords = np.array(state_to_coords([0, np.pi / 2, np.pi / 2])[:, -1])
reward = max(np.linalg.norm(final_node_coords - goal_coords), 0.0001)
reward = 1 / reward
done = False
#print(reward)
vel_reward = np.abs(self.state[4] * 5) # minus points - we dont want it to spin super fast
reward -= vel_reward
#print(vel_reward)
normalized_angle_1 = np.degrees(self.state[1])
normalized_angle_2 = np.degrees(self.state[2])
if normalized_angle_2 < 85 or normalized_angle_2 > 95:
reward -= 100
done = True
if self.state[0] < 2 and self.state[0] > -2:
pass
else:
reward -= 100
done = True
return reward, done
def step(self, action):
"""
observation - [x,phi,theta,dx,dphi,dtheta]
Num Observation Min Max
0 Cart Position -5 m 5 m
1 Pole1 Angle -pi +pi
2 Pole2 Angle -pi +pi
3 Cart Velocity -Inf Inf
4 Pole1 Angular Velocity -Inf Inf
5 Pole1 Angular Velocity -Inf Inf
"""
done = False
info = {}
self._take_action(action)
self.state_history.append(self.state)
self.action_history.append(action)
reward, done = self._reward_function()
return np.array(self.state), reward, done, info
def animate(self, i, line):
"""perform animation step"""
XY = state_to_coords(self.state_history[i])
line.set_data(XY[0], XY[1])
return line,
def render(self):
"""
Compute the render frames as specified by render_mode attribute during initialization of the environment.
"""
fig = plt.figure()
ax = fig.add_subplot(111, aspect='equal', autoscale_on=False, xlim=(-4, 4), ylim=(-2, 2))
ax.grid()
line, = ax.plot([], [], 'o-', lw=2)
animation_func = lambda i: self.animate(i, line)
ani = animation.FuncAnimation(fig, animation_func, frames=len(self.state_history), interval=20, blit=True)
return ani
def reset(self):
"""
Resets the environment to an initial state and returns the initial observation.
"""
self.state = self.init_state
self.action_history = []
self.state[1] = np.pi / 2 + np.random.uniform(-self.max_initial_angle, self.max_initial_angle)
self.state[2] = np.pi / 2 + np.random.uniform(-self.max_initial_angle, self.max_initial_angle)
self.state_history = [self.init_state]
done = False
return np.array(self.state), done