realm-framework/rf_pendulum_gemma2 at main · fracarfer5/realm-framework · 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
Generation 0
import numpy as np

def reward(action, obs):
    angle, angular_velocity = obs
    deviation = np.abs(angle - np.pi)
    # Reward based on both deviation and angular velocity
    rew = -(deviation + 0.1 * abs(angular_velocity))
    return rew

Generation 1
import numpy as np

def reward(action, obs):
    angle, angular_velocity = obs
    deviation = np.abs(angle - np.pi)
    # Reward based on both deviation and angular velocity with scaling and sparsity
    rew = -(deviation  + 0.1 * abs(angular_velocity))
    if deviation < 0.1: # Sparse reward for small deviations
        rew += 1
    return rew

Generation 2
import numpy as np

def reward(action, obs):
    angle, angular_velocity = obs
    deviation = np.abs(angle - np.pi)
    # Reward based on both deviation and angular velocity with scaling and sparsity
    rew = -(deviation  + 0.1 * abs(angular_velocity))
    if deviation < 0.05: # Sparse reward for smaller deviations, encouraging finer control
        rew += 1

    # Bonus for reaching upright position
    if deviation < 0.01:
        rew += 5
    return rew

Generation 3
import numpy as np

def reward(action, obs):
    angle, angular_velocity = obs
    deviation = np.abs(angle - np.pi)
    # Reward based on both deviation and angular velocity with scaling
    rew = - (deviation + 0.1 * abs(angular_velocity))
    if deviation < 0.05: # Sparse reward for smaller deviations, encouraging finer control
        rew += 1

    # Exponential bonus for reaching upright position to encourage faster convergence
    if deviation < 0.01:
      rew += np.exp(-deviation)
    return rew

Generation 4
import numpy as np

def reward(action, obs):
    angle, angular_velocity = obs
    deviation = np.abs(angle - np.pi)

    # Penalize high angular velocity more heavily when near upright position
    if deviation < 0.1:
      rew = -(deviation + 0.5 * abs(angular_velocity))
    else:
      rew = -(deviation + 0.1 * abs(angular_velocity))

    #  Sparse reward for smaller deviations, encouraging finer control
    if deviation < 0.05:
        rew += 1

    # Exponential bonus scaled by angular velocity for reaching upright position to encourage smoother and faster convergence
    if deviation < 0.01:
      rew += np.exp(-deviation) * (1 - abs(angular_velocity)/(12*np.pi))
    return rew