PyCMA

In [ ]:

    

%matplotlib inline

import cma
import pandas as pd
import seaborn as sns
import matplotlib
import matplotlib.pyplot as plt


    

In [ ]:

    

sns.set_context("talk")
plt.rcParams['figure.figsize'] = 20,10


    

In [ ]:

    

init_lower_bound = -10.
init_higher_bound = 10.

NUM_RUNS = 1000
DIMS = (2, 3, 4, 5, 6)
F_TARGET = 1e-30

fitness = cma.ff.sphere
fitness_name = "Sphere"

fitness = cma.ff.rastrigin
fitness_name = "Rastrigin"

fitness = cma.ff.rosen
fitness_name = "Rosenbrock"


    

In [ ]:

    

solutions_list = []
run_id = 0
num_dim = None
dim_index_list = []

def callback_function(es):
    global gen_id
    global eval_id
    global num_dim
    
    gen_id += 1
    
    indiv_dict = {
        "gen": gen_id,
        "run": run_id,
        "num_dim": num_dim,
        #"best_x": es.best.x,
        "best_f": es.best.f,
        "best_evals": es.best.evals,
        "sigma": es.sigma,
    }
    
    solutions_list.append(indiv_dict)

for ndim in DIMS:
    for run in range(1, NUM_RUNS):
        run_id = run
        num_dim = ndim
        gen_id = 0
        
        print("{}D run {}".format(ndim, run))

        x0 = np.random.uniform(low=init_lower_bound, high=init_higher_bound, size=num_dim)    # initial solution
        sigma0 = np.random.uniform(0., 1.)                                                    # initial standard deviation

        x_optimal, es = cma.fmin2(fitness, x0, sigma0,
                                  callback=callback_function,
                                  options={'verbose': -9, 'ftarget': F_TARGET, 'maxfevals': 8000})    # See last example of http://cma.gforge.inria.fr/apidocs-pycma/cma.evolution_strategy.html#fmin

df = pd.DataFrame(solutions_list)


    

In [ ]:

    

ALPHA = 0.2

ax = sns.lineplot(x="gen", y="best_f",
                  estimator=None, units="run", alpha=ALPHA,
                  data=df.loc[df.num_dim == DIMS[0]])

for dim in DIMS[1:]:
    sns.lineplot(x="gen", y="best_f",
                estimator=None, units="run", alpha=ALPHA,
                data=df.loc[df.num_dim == dim],
                ax=ax)

ax.set_yscale("log")

plt.grid()

plt.title("Best fitness ({} runs)".format(NUM_RUNS))

plt.ylabel("Best fitness")
plt.xlabel("Number of generations")

legend_lines = []
legend_str = []
for dim in DIMS:
    legend_lines.append(ax.get_lines()[(dim-1) * (NUM_RUNS-1) - 1])
    legend_str.append("{} {}D".format(fitness_name, dim))
plt.legend(legend_lines, legend_str)

plt.savefig(fitness_name.lower() + ".png");


    

In [ ]:

    

ax = sns.lineplot(x="gen", y="best_f",
                  #estimator=None, units="run", alpha=0.5,
                  data=df.loc[df.num_dim == DIMS[0]],
                  label="{} {}D".format(fitness_name, DIMS[0]))

for dim in DIMS[1:]:
    sns.lineplot(x="gen", y="best_f",
                 #estimator=None, units="run", alpha=0.5,
                 data=df.loc[df.num_dim == dim],
                 label="{} {}D".format(fitness_name, dim))

ax.set_yscale("log")

plt.grid()

plt.title("Mean of best fitness")

plt.ylabel("Mean of best fitness")
plt.xlabel("Number of generations")

plt.savefig(fitness_name.lower() + "_avg.png");


    

In [ ]:

    

ax = sns.lineplot(x="gen", y="best_f",
                  estimator=np.median,
                  data=df.loc[df.num_dim == DIMS[0]],
                  label="{} {}D".format(fitness_name, DIMS[0]))

for dim in DIMS[1:]:
    sns.lineplot(x="gen", y="best_f",
                 estimator=np.median,
                 data=df.loc[df.num_dim == dim],
                 label= "{} {}D".format(fitness_name, dim))

ax.set_yscale("log")

plt.grid()

plt.title("Median of best fitness")

plt.ylabel("Median of best fitness")
plt.xlabel("Number of generations")

plt.savefig(fitness_name.lower() + "_med.png");


    

In [ ]:

    

df.to_csv(fitness_name.lower() + ".csv")