ReinforcementLearningWithGA/ga.py at master · aralab-unr/ReinforcementLearningWithGA · 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
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
from mchgenalg import GeneticAlgorithm
import mchgenalg
import numpy as np
import os

timesEvaluated = 0
bestepochs = -1

# First, define function that will be used to evaluate the fitness
def fitness_function(genome):

    global timesEvaluated
    timesEvaluated += 1

    print("Fitness function invoked "+str(timesEvaluated)+" times")

    #setting parameter values using genome
    polyak = decode_function(genome[0:10])
    if polyak > 1:
        polyak = 1
    gamma = decode_function(genome[11:21])
    if gamma > 1:
        gamma = 1
    Q_lr = decode_function(genome[22:33])
    if Q_lr > 1:
        Q_lr = 1
    pi_lr = decode_function(genome[34:44])
    if pi_lr > 1:
        pi_lr = 1
    random_eps = decode_function(genome[45:55])
    if random_eps > 1:
        random_eps = 1
    noise_eps = decode_function(genome[56:66])
    if noise_eps > 1:
        noise_eps = 1
    epochs_default = 50 #80
    env = 'FetchSlide-v1'
    logdir ='/tmp/openaiGA'
    num_cpu = 4

    query = "python3 -m baselines.her.experiment.train --env="+env+" --logdir="+logdir+" --n_epochs="+str(epochs_default)+" --num_cpu="+str(num_cpu) + " --polyak_value="+ str(polyak) + " --gamma_value=" + str(gamma) + " --q_learning=" + str(Q_lr) + " --pi_learning=" + str(pi_lr) + " --random_epsilon=" + str(random_eps) + " --noise_epsilon=" + str(noise_eps)

    print(query)
    #calling training to calculate number of epochs required to reach close to maximum success rate
    os.system(query)
    #epochs = train.launch(env, logdir, epochs_default, num_cpu, 0, 'future', 5, 1, polyak, gamma)
    #env, logdir, n_epochs, num_cpu, seed, replay_strategy, policy_save_interval, clip_return

    file = open('epochs.txt', 'r')

    #one run is expected to converge before epochs_efault
    #if it does not converge, either add condition here, or make number of epochs as dynamic

    epochs = int(file.read())

    if epochs == None:
        epochs = epochs_default

    global bestepochs
    if bestepochs == -1:
        bestepochs = epochs
    if epochs < bestepochs:
        bestepochs = epochs
        with open('BestParameters.txt', 'a') as output:
            output.write("Epochs taken to converge : " + str(bestepochs) + "\n")
            output.write("Tau = " + str(polyak) + "\n")
            output.write("Gamma = " + str(gamma) + "\n")
            output.write("Q_learning = " + str(Q_lr) + "\n")
            output.write("pi_learning = " + str(pi_lr) + "\n")
            output.write("random_epsilon = " + str(random_eps) + "\n")
            output.write("noise_epsilon = " + str(noise_eps) + "\n")
            output.write("\n")
            output.write("=================================================")
            output.write("\n")

    print('EPOCHS taken to converge:' + str(epochs))

    print("Best epochs so far : "+str(bestepochs))
    return 1/epochs

def decode_function(genome_partial):

    prod = 0
    for i,e in reversed(list(enumerate(genome_partial))):
        if e == False:
            prod += 0
        else:
            prod += 2**abs(i-len(genome_partial)+1)
    return prod/1000

# Configure the algorithm:
population_size = 30
genome_length = 66
ga = GeneticAlgorithm(fitness_function)
ga.generate_binary_population(size=population_size, genome_length=genome_length)

# How many pairs of individuals should be picked to mate
ga.number_of_pairs = 5

# Selective pressure from interval [1.0, 2.0]
# the lower value, the less will the fitness play role
ga.selective_pressure = 1.5
ga.mutation_rate = 0.1

# If two parents have the same genotype, ignore them and generate TWO random parents
# This helps preventing premature convergence
ga.allow_random_parent = True # default True
# Use single point crossover instead of uniform crossover
ga.single_point_cross_over = False # default False

# Run 100 iteration of the algorithm
# You can call the method several times and adjust some parameters
# (e.g. number_of_pairs, selective_pressure, mutation_rate,
# allow_random_parent, single_point_cross_over)
ga.run(30) # default 1000

best_genome, best_fitness = ga.get_best_genome()

print("BEST CHROMOSOME IS")
print(best_genome)
print("It's decoded value is")
print("Tau = " + str(decode_function(best_genome[0:10])))
print("Gamma = " + str(decode_function(best_genome[11:22])))
print("Q_learning = " + str(decode_function(best_genome[23:33])))
print("pi_learning = " + str(decode_function(best_genome[34:44])))
print("random_epsilon = " + str(decode_function(best_genome[45:55])))
print("noise_epsilon = " + str(decode_function(best_genome[56:66])))

# If you want, you can have a look at the population:
population = ga.population

# and the fitness of each element:
fitness_vector = ga.get_fitness_vector()