_repeat_cosyne.py

"""Repeat the following paper for `COSYNE`:
    Gomez, F., Schmidhuber, J. and Miikkulainen, R., 2008.
    Accelerated neural evolution through cooperatively coevolved synapses.
    Journal of Machine Learning Research, 9(31), pp.937-965.
    https://jmlr.org/papers/v9/gomez08a.html

    We notice that the EvoTorch library (recently developed in the company of two original authors)
      provides an open-source implementation for it:
    https://docs.evotorch.ai/v0.3.0/reference/evotorch/algorithms/ga/#evotorch.algorithms.ga.Cosyne

    In our implementation, we refer to this open-source implementation rather than the original paper,
    but with a slight simplification for the *permutation* operator.

    Luckily our Python code could repeat the data reported in the reference library *well*.
    Therefore, we argue that its repeatability could be **well-documented**.



    The reference Python code based on EvoTorch is given below:
    -----------------------------------------------------------
    import torch
    from evotorch import Problem
    from evotorch.algorithms import ga
    from evotorch.logging import StdOutLogger


    def norm(x: torch.Tensor) -> torch.Tensor:
        return (torch.linalg.norm(x, dim=-1)).pow(2)


    problem = Problem('min', norm, initial_bounds=(-5.0, 5.0), solution_length=1000)
    searcher = ga.Cosyne(problem, popsize=100, tournament_size=10, mutation_probability=1.0, mutation_stdev=1.0)
    logger = StdOutLogger(searcher)
    searcher.run(num_generations=3000)
"""
import time

import numpy as np

from pypop7.benchmarks.base_functions import sphere
from pypop7.optimizers.cc.cosyne import COSYNE as Solver


if __name__ == '__main__':
    start_run = time.time()
    ndim_problem = 1000
    for f in [sphere]:
        print('*' * 7 + ' ' + f.__name__ + ' ' + '*' * 7)
        problem = {'fitness_function': f,
                   'ndim_problem': ndim_problem,
                   'lower_boundary': -5*np.ones((ndim_problem,)),
                   'upper_boundary': 5*np.ones((ndim_problem,))}
        options = {'max_function_evaluations': 150*3000 + 100,
                   'sigma': 1.0,
                   'seed_rng': 0}
        solver = Solver(problem, options)
        results = solver.optimize()
        print(results)  # 7368.388082444299 vs 7388.05517578125 (from EvoTorch library)
        print('*** Runtime: {:7.5e}'.format(time.time() - start_run))