import torch
from evox.core import Algorithm, Mutable, Parameter
from evox.utils import clamp
from .utils import min_by
[docs]
class SLPSOGS(Algorithm):
"""The basic Particle Swarm Optimization Social Learning PSO Using Gaussian Sampling for Demonstrator Choice (SLPSOGS) algorithm.
## Class Methods
* `__init__`: Initializes the SLPSOGS algorithm with given parameters (population size, inertia weight, cognitive weight, and social weight).
* `setup`: Initializes the SLPSOGS algorithm with given lower and upper bounds for particle positions, and sets up initial population, velocity, and buffers for tracking best local and global positions and fitness values.
* `step`: Performs a single optimization step using Particle Swarm Optimization (SLPSOGS), updating local best positions and fitness values, and adjusting velocity and positions based on inertia, cognitive, and social components.
Note that the `evaluate` method is not defined in this class, it is a proxy function of `Problem.evaluate` set by workflow; therefore, it cannot be used in class methods other than `step`.
"""
def __init__(
self,
pop_size: int,
lb: torch.Tensor,
ub: torch.Tensor,
social_influence_factor: float = 0.2, # epsilon
demonstrator_choice_factor: float = 0.7, # theta
device: torch.device | None = None,
):
"""
Initialize the SLPSOGS algorithm with the given parameters.
:param pop_size: The size of the population.
:param lb: The lower bounds of the particle positions. Must be a 1D tensor.
:param ub: The upper bounds of the particle positions. Must be a 1D tensor.
:param w: The inertia weight. Defaults to 0.6.
:param phi_p: The cognitive weight. Defaults to 2.5.
:param phi_g: The social weight. Defaults to 0.8.
:param device: The device to use for the tensors. Defaults to None.
"""
super().__init__()
device = torch.get_default_device() if device is None else device
assert lb.shape == ub.shape and lb.ndim == 1 and ub.ndim == 1 and lb.dtype == ub.dtype
self.dim = lb.shape[0]
self.pop_size = pop_size
# Here, Parameter is used to indicate that these values are hyper-parameters
# so that they can be correctly traced and vector-mapped
self.social_influence_factor = Parameter(social_influence_factor, device=device)
self.demonstrator_choice_factor = Parameter(demonstrator_choice_factor, device=device)
# setup
lb = lb[None, :].to(device=device)
ub = ub[None, :].to(device=device)
length = ub - lb
pop = torch.rand(self.pop_size, self.dim, device=device)
pop = length * pop + lb
velocity = torch.rand(self.pop_size, self.dim, device=device)
velocity = 2 * length * velocity - length
# write to self
self.lb = lb
self.ub = ub
# mutable
self.pop = Mutable(pop)
self.fit = Mutable(torch.empty(self.pop_size, device=device))
self.velocity = Mutable(velocity)
self.global_best_location = Mutable(pop[0])
self.global_best_fit = Mutable(torch.tensor(torch.inf, device=device))
[docs]
def init_step(self):
self.fit = self.evaluate(self.pop)
self.global_best_fit = torch.min(self.fit)
[docs]
def step(self):
"""Perform a normal optimization step using SLPSOGS."""
device = self.pop.device
global_best_location, global_best_fit = min_by(
[self.global_best_location.unsqueeze(0), self.pop],
[self.global_best_fit.unsqueeze(0), self.fit],
)
# Demonstrator Choice
# sort from largest fitness to smallest fitness (worst to best)
ranked_population = self.pop[torch.argsort(self.fit, descending=True)]
sigma = self.demonstrator_choice_factor * (self.pop_size - (torch.arange(self.pop_size, device=device) + 1))
# normal distribution (shape=(self.pop_size,)) means
# each individual choose a demonstrator by normal distribution
# with mean = pop_size and std = sigma
standard_normal_distribution = torch.randn(self.pop_size, device=device)
normal_distribution = sigma * (-torch.abs(standard_normal_distribution)) + self.pop_size
index_k = torch.clamp(normal_distribution, 1, self.pop_size).to(dtype=torch.int64) - 1
X_k = ranked_population[index_k]
# Update population and velocity
X_avg = torch.mean(self.pop, dim=0)
r1 = torch.rand(self.pop_size, self.dim, device=device)
r2 = torch.rand(self.pop_size, self.dim, device=device)
r3 = torch.rand(self.pop_size, self.dim, device=device)
velocity = r1 * self.velocity + r2 * (X_k - self.pop) + r3 * self.social_influence_factor * (X_avg - self.pop)
pop = self.pop + velocity
pop = clamp(pop, self.lb, self.ub)
velocity = clamp(velocity, self.lb, self.ub)
self.pop = pop
self.velocity = velocity
self.global_best_location = global_best_location
self.global_best_fit = global_best_fit
self.fit = self.evaluate(self.pop)
EvoX