-
Notifications
You must be signed in to change notification settings - Fork 0
/
single_intersection_evaluation.py
174 lines (143 loc) · 6.2 KB
/
single_intersection_evaluation.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
from distutils.log import info
import os
from pyexpat import model
import sys
if 'SUMO_HOME' in os.environ:
tools = os.path.join(os.environ['SUMO_HOME'], 'tools')
sys.path.append(tools)
else:
sys.exit("Please declare the environment variable 'SUMO_HOME'")
from sumo_rl import SumoEnvironment
import traci
from stable_baselines3 import A2C, PPO, DQN
from stable_baselines3.common.utils import obs_as_tensor, get_device
from collections import Counter
import numpy as np
import torch
def predict_proba(model, state):
"""
Takes as input a model and an observation. It returns the probability distribution for each action.
Works only on PPO and A2C algorithms.
"""
obs = obs_as_tensor(state.reshape([1,-1]), model.policy.device)
dis = model.policy.get_distribution(obs)
probs = dis.distribution.probs
probs_np = probs.cpu().detach().numpy()
return probs_np[0]
def action_probability(model, state):
"""
Takes in a model and an observation. It outputs the probability for each action.
It works for DQN.
"""
action_predicted = model.predict(state)[0]
exploration_eps = model.exploration_rate
exploitation_eps = 1 - exploration_eps
probs_np = [exploration_eps/4 for i in range(4)]
probs_np[action_predicted] = exploitation_eps + exploration_eps/4
return np.array(probs_np)
def majority_vote(p1, p2, p3=[]):
"""
Receives as input three actions and outputs the action that receives the most votes.
In case all action receive 1 vote it outputs a random action.
"""
l = [np.argmax(p) for p in [p1, p2, p3] if len(p) > 1]
if len(set(l)) < 3:
return max(set(l), key=l.count)
else:
return rank_voting(p1, p2, p3)
def majority_vote_1(p1, p2, p3):
"""
Receives as input three actions and outputs the action that receives the most votes.
In case all action receive 1 vote it outputs a random action.
"""
l = [np.argmax(p) for p in [p1, p2, p3] if len(p) > 1]
if len(set(l)) < 3:
return max(set(l), key=l.count)
else:
return np.random.choice(a=l, p=[.1, .75, .15])
def majority_vote_with_probs(p1, p2, p3=[]):
'''
Takes three different probability vectors in and outputs a randomly sampled
action from n_action according to majority voting scheme
From: https://medium.com/@tu_53768/ensemble-reinforcement-learning-b06e28eec31c#:~:text=from%20these%20vectors.-,Majority%20vote,-First%2C%20the%20majority
'''
a = range(4)
a1 = np.random.choice(a=a, p=p1)
a2 = np.random.choice(a=a, p=p2)
if len(p3) > 1:
a3 = np.random.choice(a=a, p=p3)
l = [a1, a2, a3]
else:
l = [a1, a2]
return max(set(l), key=l.count)
def soft_voting(p1, p2, p3=[], probs=False):
if len(p3) > 1:
soft_vote = np.sum(np.stack((p1, p2, p3)), axis=0)
else:
soft_vote = np.sum(np.stack((p1, p2)), axis=0)
if not probs:
return np.argmax(soft_vote)
else:
return soft_vote
def rank_voting_1(p1, p2, p3=np.zeros(4), probs=False):
p = np.stack((p1, p2, p3), axis=0)
max_probs = np.max(p, axis=0)
best_action = np.argmax(max_probs)
if not probs:
return best_action
else:
return max_probs
def rank_voting(p1, p2, p3=[], probs=False):
if len(p3) > 1:
for p in [p1, p2, p3]:
p_copy = np.sort(p)
for i in range(len(p)):
p[np.where(p==p_copy[i])] = (i+1)*p_copy[i]
soft_vote = np.sum(np.stack((p1, p2, p3)), axis=0)
else:
soft_vote = np.sum(np.stack((p1, p2)), axis=0)
if not probs:
return np.argmax(soft_vote)
else:
return soft_vote
def average_voting(p1, p2, p3=[], T=0.5):
'''
Takes in different probability vectors and outputs a randomly sampled
action from n_action with probability equals the average probability of the
normalized exponentiated input vectors, with a temperature T controlling
the degree of spread for the out vector (Boltzmann Probabilities)
'''
a = range(4)
if len(p3) > 1:
boltz_ps = [np.exp(prob/T)/sum(np.exp(prob/T)) for prob in [p1, p2, p3]]
p = (boltz_ps[0] + boltz_ps[1] + boltz_ps[2])/3
p /= np.sum(p) # To avoid (ValueError: np-random-choice-probabilities-do-not-sum-to-1)[https://stackoverflow.com/questions/46539431/np-random-choice-probabilities-do-not-sum-to-1]
a = np.random.choice(a=a, p=p)
else:
boltz_ps = [np.exp(prob/T)/sum(np.exp(prob/T)) for prob in [p1, p2]]
p = (boltz_ps[0] + boltz_ps[1])/2
p /= np.sum(p) # To avoid (ValueError: np-random-choice-probabilities-do-not-sum-to-1)[https://stackoverflow.com/questions/46539431/np-random-choice-probabilities-do-not-sum-to-1]
a = np.random.choice(a=a, p=p)
return a
if __name__ == '__main__':
env = SumoEnvironment(net_file='/home/talos/MSc_Thesis/nets/2way-single-intersection/single-intersection.net.xml',
route_file='/home/talos/MSc_Thesis/nets/2way-single-intersection/single-intersection-vhvh.rou.xml',
# out_csv_name='/home/talos/MSc_Thesis/outputs/2way-single-intersection/A2C-trained-5',
single_agent=True,
use_gui=True,
num_seconds=100000,
sumo_seed=42)
model_dqn = DQN.load('/home/talos/MSc_Thesis/models/DQN-06012022 171843/DQN2M') # Best performance on predictions DQN-trained-4
model_ppo = PPO.load('/home/talos/MSc_Thesis/models/PPO-05122022 164805/101000') # Best performance on predictions PPO-trained-3
model_a2c = A2C.load('/home/talos/MSc_Thesis/models/A2C-05312022 165157/A2C1.5M') # Best performance on predictions A2C-trained-0
# model = A2C.load('/home/talos/MSc_Thesis/models/A2C-06062022 082522/A2C100')
EPISODES = 1
for ep in range(EPISODES):
obs = env.reset()
done = False
while not done:
env.render()
action = rank_voting(action_probability(model_dqn, obs), predict_proba(model_a2c, obs))
# action, _ = model_ppo.predict(obs)
obs, reward, done, info = env.step(action)
env.close()