from typing import Literal
import torch
from evox.core import Algorithm, Mutable, Parameter
from .adam_step import adam_single_tensor
[docs]
class GuidedES(Algorithm):
"""The implementation of the Guided-ES algorithm.
Reference:
Guided evolutionary strategies: Augmenting random search with surrogate gradients
(https://arxiv.org/abs/1806.10230)
This code has been inspired by or utilizes the algorithmic implementation from evosax.
More information about evosax can be found at the following URL:
GitHub Link: https://github.com/RobertTLange/evosax
"""
def __init__(
self,
pop_size: int,
center_init: torch.Tensor,
subspace_dims: int | None = None,
optimizer: Literal["adam"] | None = None,
sigma: float = 0.03,
lr: float = 60,
sigma_decay: float = 1.0,
sigma_limit: float = 0.01,
device: torch.device | None = None,
):
"""Initialize the Guided-ES algorithm with the given parameters.
:param pop_size: The size of the population.
:param center_init: The initial center of the population. Must be a 1D tensor.
:param optimizer: The optimizer to use. Defaults to None. Currently, only "adam" or None is supported.
:param lr: The learning rate for the optimizer. Defaults to 0.05.
:param sigma: The standard deviation of the noise. Defaults to 0.03.
:param sigma_decay: The decay factor for the standard deviation. Defaults to 1.0.
:param sigma_limit: The minimum value for the standard deviation. Defaults to 0.01.
:param subspace_dims: The dimension of the subspace. Defaults to None.
:param device: The device to use for the tensors. Defaults to None.
"""
super().__init__()
assert pop_size > 1 and pop_size % 2 == 0
dim = center_init.shape[0]
if subspace_dims is None:
subspace_dims = dim
# set hyperparameters
self.beta = Parameter(1.0, device=device)
self.lr = Parameter(lr, device=device)
self.sigma_decay = Parameter(sigma_decay, device=device)
self.sigma_limit = Parameter(sigma_limit, device=device)
# set value
self.dim = dim
self.pop_size = pop_size
self.optimizer = optimizer
self.subspace_dims = subspace_dims
# setup
center_init = center_init.to(device=device)
self.center = Mutable(center_init)
self.alpha = Mutable(torch.tensor(0.5, device=device))
self.sigma = Mutable(torch.tensor(sigma, device=device))
self.grad_subspace = Mutable(torch.randn(subspace_dims, dim, device=device))
if optimizer == "adam":
self.exp_avg = Mutable(torch.zeros_like(self.center))
self.exp_avg_sq = Mutable(torch.zeros_like(self.center))
self.beta1 = Parameter(0.9, device=device)
self.beta2 = Parameter(0.999, device=device)
[docs]
def step(self):
"""Run one step of the Guided-ES algorithm.
The function will sample a population, evaluate their fitness, and then
update the center and standard deviation of the algorithm using the
sampled population.
"""
device = self.center.device
a = self.sigma * torch.sqrt(self.alpha / self.dim)
c = self.sigma * torch.sqrt((1.0 - self.alpha) / self.subspace_dims)
eps_full = torch.randn(self.dim, int(self.pop_size // 2), device=device)
eps_subspace = torch.randn(self.subspace_dims, int(self.pop_size // 2), device=device)
Q, _ = torch.linalg.qr(self.grad_subspace)
z_plus = a * eps_full + c * (Q @ eps_subspace)
z_plus = torch.swapaxes(z_plus, 0, 1)
z = torch.cat([z_plus, -1.0 * z_plus])
population = self.center + z
fitness = self.evaluate(population)
noise = z / self.sigma
noise_1 = noise[: int(self.pop_size / 2)]
fit_1 = fitness[: int(self.pop_size / 2)]
fit_2 = fitness[int(self.pop_size / 2) :]
fit_diff = fit_1 - fit_2
fit_diff_noise = noise_1.T @ fit_diff
theta_grad = (self.beta / self.pop_size) * fit_diff_noise
self.grad_subspace = torch.cat([self.grad_subspace, theta_grad[None, :]])[1:, :]
if self.optimizer is None:
center = self.center - self.lr * theta_grad
else:
center, self.exp_avg, self.exp_avg_sq = adam_single_tensor(
self.center,
theta_grad,
self.exp_avg,
self.exp_avg_sq,
self.beta1,
self.beta2,
self.lr,
)
self.center = center
sigma = torch.maximum(self.sigma_decay * self.sigma, self.sigma_limit)
self.sigma = sigma
[docs]
def record_step(self):
return {"center": self.center, "sigma": self.sigma}
EvoX