import torch
[docs]
def simulated_binary(x: torch.Tensor, pro_c: float = 1.0, dis_c: float = 20.0) -> torch.Tensor:
"""Simulated binary crossover (SBX)
:param x: Parent solutions in a 2D tensor (size: n x d).
:param pro_c: Probability of doing crossover.
:param dis_c: Distribution index of SBX.
:return: Offspring solutions after crossover.
"""
n, m = x.size()
parent1_dec = x[: n // 2, :]
parent2_dec = x[n // 2 : n // 2 * 2, :]
# Uniform distribution for mutation
mu = torch.rand(n // 2, m, device=x.device)
# Beta calculation for SBX
beta = torch.zeros(mu.size(), device=x.device)
beta = torch.where(mu <= 0.5, torch.pow(2 * mu, 1 / (dis_c + 1)), beta)
beta = torch.where(mu > 0.5, torch.pow(2 - 2 * mu, -1 / (dis_c + 1)), beta)
# Random binary for mutation direction
beta = beta * (1 - torch.randint(0, 2, beta.size(), device=x.device) * 2)
# Apply crossover probability to mutate
beta = torch.where(torch.rand(mu.size(), device=x.device) < 0.5, 1, beta)
beta = torch.where(torch.rand(mu.size(), device=x.device) > pro_c, 1, beta)
offspring_dec = torch.cat(
[
(parent1_dec + parent2_dec) / 2 + beta * (parent1_dec - parent2_dec) / 2,
(parent1_dec + parent2_dec) / 2 - beta * (parent1_dec - parent2_dec) / 2,
]
)
return offspring_dec
EvoX