Introduction

pycomocma is a Python implementation of COMO-CMA-ES which is a Multiobjective Evolution Strategy, based upon the Covariance Matrix Adaptation Evolution Strategy (CMA-ES) single optimizer.

For the time being, only the bi-objective case is tested and functional.

Installation

Either via

pip install git+https://github.com/CMA-ES/pycomocma.git@master

or simply via

pip install comocma

Links

• Code on Github
• Documentation in
  • apidocs format
  • epydocs format

Testing of the comocma module

The script

python -m comocma

runs the test written in the __main__ file.

Use cases

Instantiating a multiobjective solver

Importing necessary packages:

import cma, comocma

Setting parameters:

dimension = 10  # dimension of the search space
num_kernels = 5 # number of single-objective solvers (number of points on the front)
sigma0 = 0.2    # initial step-sizes

Instantiate a multiobjective solver

list_of_solvers = comocma.get_cmas(num_kernels * [dimension * [0]], sigma0) # produce `num_kernels cma instances`
moes = comocma.Sofomore(list_of_solvers, reference_point=[11, 11]) # create a bi-objective como-cma-es instance
moes3 = comocma.Sofomore(list_of_solvers, reference_point=[11, 11, 11]) # create a multiobjective como-cma-es instance

Setting a callable multiobjective function

fitness = comocma.FitFun(cma.ff.sphere, lambda x: cma.ff.sphere(x-1)) # a callable bi-objective function
fitness3 = comocma.FitFun(cma.ff.sphere, lambda x: cma.ff.sphere(x-1), lambda x: cma.ff.sphere(x+1)) # a callable multiobjective function

Single-objective options: a use case with few cma-es' options

list_of_solvers = comocma.get_cmas(num_kernels * [dimension * [0]], 0.2, inopts={'bounds': [0.2, 0.9], 'tolx': 10**-7,'popsize': 32}) 
# produce `num_kernels cma instances`
moes = comocma.Sofomore(list_of_solvers, [1.1, 1.1]) # create a como-cma-es instance

Use case with some Multiobjective options

list_of_solvers = comocma.get_cmas(num_kernels * [dimension * [0]], 0.2)
moes = comocma.Sofomore(list_of_solvers, [1.1, 1.1], opts={'archive': True, 'restart': None, 'update_order': None}) # create a como-cma-es instance

The Optimize interface

Initialization

import cma, comocma

dimension = 10  # dimension of the search space
num_kernels = 5 # number of single-objective solvers (number of points on the front)
sigma0 = 0.2    # initial step-sizes

list_of_solvers = comocma.get_cmas(num_kernels * [dimension * [0]], sigma0) # produce `num_kernels cma instances`
moes = comocma.Sofomore(list_of_solvers, [11,11]) # create a como-cma-es instance

fitness = comocma.FitFun(cma.ff.sphere, lambda x: cma.ff.sphere(x-1)) # a callable bi-objective function

Optimizing fitness until default stopping criteria

moes.optimize(fitness)

Iterat #Fevals   Hypervolume   axis ratios   sigmas   min&max stds
                                  (median)  (median)    (median)
    1     10 1.210000000000000e+00 1.0e+00 2.00e-01  2e-01  2e-01
    2     20 1.210000000000000e+00 1.0e+00 2.00e-01  2e-01  2e-01
    3     30 1.210000000000000e+00 1.0e+00 1.85e-01  2e-01  2e-01
  100   1000 1.207601015381810e+00 1.6e+00 3.40e-02  3e-02  3e-02
  200   2000 1.209903687756354e+00 1.7e+00 7.74e-03  5e-03  6e-03
  300   3000 1.209997694077156e+00 1.8e+00 2.03e-03  1e-03  1e-03
  400   4000 1.209999800600613e+00 1.8e+00 4.90e-04  2e-04  3e-04
  480   4800 1.209999979594839e+00 1.9e+00 2.02e-04  7e-05  9e-05

Optimizing fitness with a limited number of iterations

moes.optimize(fitness, iterations=300)

Iterat #Fevals   Hypervolume   axis ratios   sigmas   min&max stds
                                (median)  (median)    (median)
  1     10 1.100000000000000e+01 1.0e+00 2.00e-01  2e-01  2e-01
  2     20 2.158412269365152e+01 1.0e+00 2.00e-01  2e-01  2e-01
  3     30 2.896035267829712e+01 1.0e+00 1.98e-01  2e-01  2e-01
100   1000 9.512982413314423e+01 1.7e+00 1.01e-01  8e-02  9e-02
200   2000 9.703624875547615e+01 1.9e+00 4.27e-02  3e-02  4e-02
300   3000 9.722958234416403e+01 1.9e+00 1.63e-02  9e-03  1e-02

Optimizing fitness with a maximum number of evaluations

moes.optimize(fitness, maxfun=3000)

Iterat #Fevals   Hypervolume   axis ratios   sigmas   min&max stds
                                (median)  (median)    (median)
  1     10 1.100000000000000e+01 1.0e+00 2.00e-01  2e-01  2e-01
  2     20 2.158412269365152e+01 1.0e+00 2.00e-01  2e-01  2e-01
  3     30 2.896035267829712e+01 1.0e+00 1.98e-01  2e-01  2e-01
100   1000 9.512982413314423e+01 1.7e+00 1.01e-01  8e-02  9e-02
200   2000 9.703624875547615e+01 1.9e+00 4.27e-02  3e-02  4e-02
300   3000 9.722958234416403e+01 1.9e+00 1.63e-02  9e-03  1e-02

The ask-and-tell interface

while not moes.stop():
    solutions = moes.ask("all")
    objective_values = [fitness(x) for x in solutions]
    moes.tell(solutions, objective_values)
    moes.disp()          # display datas during the optimization
    moes.logger.add()    # logging data after each `ask` and `tell` call

Iterat #Fevals   Hypervolume   axis ratios   sigmas   min&max stds
                                  (median)  (median)    (median)
    1    180 1.990425600000000e-01 1.0e+00 1.88e-01  2e-01  2e-01
    2    360 2.279075246432772e-01 1.1e+00 1.87e-01  2e-01  2e-01
    3    540 2.436105134581627e-01 1.2e+00 1.90e-01  2e-01  2e-01
  100  18000 3.607157703968831e-01 2.1e+00 1.80e-02  1e-02  2e-02
  200  35172 3.635275131024869e-01 2.1e+00 5.95e-03  4e-03  5e-03
  300  49788 3.637412031970786e-01 2.2e+00 1.29e-03  8e-04  1e-03
  320  50784 3.637421277015990e-01 2.2e+00 1.26e-03  7e-04  9e-04

Argument of moes.ask

solutions = moes.ask() # we generate offspring for only one kernel (sequential)
solutions = moes.ask("all") # we generate offspring simultaneously for all kernels (parallel)
solutions = moes.ask(number_asks) # we generate offspring for `number_asks` kernels

Picklable object: saving and resuming a MO optimization with the ask-and-tell interface

Initialization

import cma, como, pickle

dimension = 10  # dimension of the search space
num_kernels = 5 # number of single-objective solvers (number of points on the front)
sigma0 = 0.2    # initial step-sizes

list_of_solvers = como.get_cmas(num_kernels * [dimension * [0]], sigma0) # produce `num_kernels cma instances`
moes = como.Sofomore(list_of_solvers, reference_point = [11,11]) # create a como-cma-es instance

fitness = como.FitFun(cma.ff.sphere, lambda x: cma.ff.sphere(x-1)) # a callable bi-objective function

Saving an optimization

for i in range(100):
    solutions = moes.ask()
    objective_values = [fitness(x) for x in solutions]
    moes.tell(solutions, objective_values)
    moes.disp()

pickle.dump(moes, open('saved-mocma-object.pkl', 'wb')) # we save the instance
print('saved')
del moes  # deleting completely the Sofomore instance

Output

Iterat #Fevals   Hypervolume   axis ratios   sigmas   min&max stds
                                  (median)  (median)    (median)
    1     10 1.100000000000000e+01 1.0e+00 2.00e-01  2e-01  2e-01
    2     20 2.845200549045931e+01 1.0e+00 2.00e-01  2e-01  2e-01
    3     30 3.440089785096067e+01 1.0e+00 2.00e-01  2e-01  2e-01
  100   1000 9.562953505152342e+01 1.9e+00 1.13e-01  9e-02  1e-01
saved

Resuming an optimization

moes = pickle.load(open('saved-mocma-object.pkl', 'rb')) # we load the saved file here

moes.optimize(fitness, iterations=400)

Output

200   2000 9.716644477685412e+01 1.9e+00 3.33e-02  2e-02  3e-02
300   3000 9.723550009906029e+01 2.0e+00 1.13e-02  6e-03  8e-03
400   4000 9.724067117112808e+01 1.9e+00 2.95e-03  1e-03  2e-03
500   5000 9.724107479961819e+01 2.0e+00 9.38e-04  4e-04  5e-04

Example of plots

COMO-CMA-ES data plottings

moes.logger.plot_front()

moes.logger.plot_divers()

CMA-ES plots of written data

cma.plot("cma_kernels/0")