BBOB torch

Implementation of BBOB functions (Real-Parameter Black-Box Optimization Benchmarking) as specified by https://coco.gforge.inria.fr/downloads/download16.00/bbobdocfunctions.pdf.

All the functions are vectorized and allow to pass potentional solutions in the same (num_of_solutions, problem_dimension).

Problem description

All the problems are represented by Problem class. This class allows to:

• Evaluate your solutions by directly calling it problem(solutions).
• Get problem dimension problem.dim.
• Get optimal solution resp. optimal function value using problem.x_opt resp. problem.f_opt.
• Get boundaries of solution using problem.min resp. problem.max for each dimension.
• Change underlying type or device using problem.type(torch.float16) and problem.ty(torch.device('cuda:0')).

Problem creation

You can create new instance of each problem by calling corresponding create_fxx function. This function accepts problem dimension and can optionally accept device and seed.

import torch
import bbobtorch
problem = bbobtorch.create_f09(40, dev=torch.device('cuda:0'), seed=42)

Example

import matplotlib.pyplot as plt
import numpy as np
import torch
import bbobtorch

x = torch.arange(-5,5, 0.01, dtype=torch.float32)
grid = torch.stack(torch.meshgrid(x, x), -1)
flat_grid = torch.reshape(grid, (-1,2))
xgrid, ygrid = np.meshgrid(x.numpy(), x.numpy())

fn = bbobtorch.create_f22(2, seed=42)  # two dimension with seed 42
results = fn(flat_grid)
results_grid = torch.reshape(results, xgrid.shape) - fn.f_opt

plt.figure(figsize=(6,6))
plt.pcolormesh(xgrid, ygrid, results_grid, cmap='inferno', shading='nearest')
plt.scatter(*fn.x_opt.tolist()[::-1], marker='x', c='r')
plt.show()

You can view all the functions in attached PDF.

Author: Patrik Valkovič

License: MIT