I've been studying about Genetic Algorithms lately and I decided to make my own using Python. I'll share the working I have done, below.
# Generates Random Population
def generate_random_population(npop, limits=list(zip(np.zeros(5),np.ones(5))), ngenes=5):
def new_ind():
return [random.uniform(limits[i][0], limits[i][1]) for i in range(ngenes)]
return np.array([new_ind() for n in range(npop)])
# Function to evaluate all individuals and give them a score
# fopt1 only has a minimum (unimodal) at x = (0,0, ..., 0) in which fopt1 = 0.
def fopt1(ind):
x0 = [ind[len(ind)-1]]
xlast = [ind[0]]
working_array = np.concatenate((x0,ind,xlast))
res = 0
for j in range(1, len(ind)+1):
res += (2*working_array[j-1] + (working_array[j]**2)*working_array[j+1] - working_array[j+1])**2
return res
# Receives a certain population of individuals and an evaluation function (usually called * fitness function *) and returns an ordered list of tuples
def eval_pop(pop, f):
# Returns a list of tuples in descending order of the goodness of f. Shape of tuples are (individual, score), e.g., ([2.3,0.004,1,8.2,6], 0.361).
list = []
for i in pop:
j = (pop, f(pop))
list.append(j)
return list
# Function to produce a next generation of individuals is to select the pairs that will interbreed to have offspring
def couples_selection(ordered_pop, n_elitism):
if len(ordered_pop) < 10:
print("Error: population's size should be higher than 9")
return
len_a = int(len(ordered_pop)/10)
len_b = len_a * 3
len_c = len_a * 4
a = np.ones(len_a) * 0.5 / len_a
b = np.ones(len_b) * 0.3 / len_b
c = np.ones(len_c) * 0.15 / len_c
d = np.ones(len(ordered_pop) - len_a*8)
d = d * 0.05 / len(d)
prob = np.concatenate((a,b,c,d))
indices = range(len(ordered_pop))
selected_indices = [choice(indices, 2, p=prob) for i in range(len(ordered_pop) - n_elitism)]
couples = [[ordered_pop[i1], ordered_pop[i2]] for [i1,i2] in selected_indices]
return np.array(couples)
def mutate(ind, limits):
# print("Mutating individual ", ind)
factor = 1 + (0.2 * choice([-1,1], 1))
gene_index = choice(range(len(ind)), 1)[0]
mutated_val = ind.item(gene_index) * factor
if mutated_val < limits[gene_index][0]:
mutated_val = limits[gene_index][0]
elif mutated_val > limits[gene_index][1]:
mutated_val = limits[gene_index][1]
ind[gene_index] = mutated_val
return
def crossover(couple):
ancestor1 = couple[0]
ancestor2 = couple[1]
c1, c2 = ancestor1.copy(), ancestor2.copy()
pt = randint(1, len(ancestor1)-2)
# perform crossover
c1 = ancestor1[:pt] + ancestor2[pt:]
c2 = ancestor2[:pt] + ancestor1[pt:]
return [c1, c2]
def get_offspring(couples, mutp, limits):
children = [crossover(couple) for couple in couples]
mutation_roulette = [choice([True, False], 1, p=[mutp, 1-mutp]) for _ in children]
children_roulette = list(zip(children, mutation_roulette))
for child in children_roulette:
if child[1][0]:
mutate(child[0], limits)
# print("Mutated: ",child[0])
return np.array([child[0] for child in children_roulette])
When I run the following driver function with the following function call:
runGA(100, 5, list(zip(np.ones(5)*-2,np.ones(5)*2)), fopt13, 4, 0.4, 25)
def runGA(npop, ngenes, limits, fitness, nelitism, mutp, ngenerations):
pop = generate_random_population(npop, limits, ngenes)
sorted_pop_with_score = eval_pop(pop, fitness)
new_pop = np.array([p[0] for p in sorted_pop_with_score])
for g in range(ngenerations):
# TO DO: Complete your GA!
couples = couples_selection(new_pop, nelitism)
popp = get_offspring(couples,mutp, limits)
eval_pop_result = eval_pop(pop,fitness)
# END OF TO DO
print("Winner after generation", g, ":", eval_pop_result[0])
print("Absolute winner:")
return sorted_pop_with_score[0]
I'm getting this error in the crossover function:
ValueError Traceback (most recent call last)
<ipython-input-20-375adbb7b149> in <module>
----> 1 runGA(100, 5, list(zip(np.ones(5)*-2,np.ones(5)*2)), fopt13, 4, 0.4, 25)
<ipython-input-12-6619b9c7d476> in runGA(npop, ngenes, limits, fitness, nelitism, mutp, ngenerations)
8 # TO DO: Complete your GA!
9 couples = couples_selection(new_pop, nelitism)
---> 10 popp = get_offspring(couples,mutp, limits)
11 eval_pop_result = eval_pop(pop,fitness)
12
<ipython-input-16-5e8ace236573> in get_offspring(couples, mutp, limits)
34 def get_offspring(couples, mutp, limits):
35
---> 36 children = [crossover(couple) for couple in couples]
37 mutation_roulette = [choice([True, False], 1, p=[mutp, 1-mutp]) for _ in children]
38 children_roulette = list(zip(children, mutation_roulette))
<ipython-input-16-5e8ace236573> in <listcomp>(.0)
34 def get_offspring(couples, mutp, limits):
35
---> 36 children = [crossover(couple) for couple in couples]
37 mutation_roulette = [choice([True, False], 1, p=[mutp, 1-mutp]) for _ in children]
38 children_roulette = list(zip(children, mutation_roulette))
<ipython-input-16-5e8ace236573> in crossover(couple)
25 print(len(ancestor1))
26 print(len(ancestor2))
---> 27 c1 = ancestor1[:pt] + ancestor2[pt:]
28 c2 = ancestor2[:pt] + ancestor1[pt:]
29
ValueError: operands could not be broadcast together with shapes (39,5) (61,5)
I also tried the np.concatenate function but it gives the following error on the same step:
TypeError: only integer scalar arrays can be converted to a scalar index
Any help would be highly appreciated!
My comments turned into an answer:
So it looks like you need to run couples_selection()
on the population for each generation, then run get_offspring()
on the couples returned from couples_selection()
, and then run eval_pop()
on the population returned from get_offspring()
. Then, the winner of that generation will be the individual from the returned list of eval_pop()
that had the highest score. It looks like eval_pop()
is supposed to sort its returned list in descending order of score, but doesn't appear to; otherwise, the [0]
index of the returned list would be the one with the highest score, aka the winner.
Also, if you're returning sorted_pop_with_score[0]
as the absolute winner, then it seems like you need to be adding the winner of each generation to some list, and then run eval_pop()
on that list after you complete all the generations, and set sorted_pop_with_score
to the result of that final eval_pop()
.