自定义算法和问题#

在这个笔记本中,我们将展示如何使用AlgorithmProblem来创建一个自定义算法和问题。这里我们将给出一个实现解决Sphere问题的PSO算法的例子。

from typing import List

import torch

from evox.core import Algorithm, Mutable, Parameter, Problem, jit_class
from evox.utils import clamp
from evox.workflows import EvalMonitor, StdWorkflow

算法示例:PSO 算法#

粒子群优化(PSO)是一种基于种群的元启发式算法,灵感来自鸟类和鱼类的社会行为。它广泛用于解决连续和离散优化问题。

以下是EvoX中PSO算法的实现示例:

@jit_class
class PSO(Algorithm):
    # Initialize the PSO algorithm with the given parameters.
    def __init__(
        self,
        pop_size: int,
        lb: torch.Tensor,
        ub: torch.Tensor,
        w: float = 0.6,
        phi_p: float = 2.5,
        phi_g: float = 0.8,
        device: torch.device | None = None,
    ):
        super().__init__()
        assert lb.shape == ub.shape and lb.ndim == 1 and ub.ndim == 1 and lb.dtype == ub.dtype
        self.pop_size = pop_size
        self.dim = lb.shape[0]
        # Here, Parameter is used to indicate that these values are hyper-parameters
        # so that they can be correctly traced and vector-mapped
        self.w = Parameter(w, device=device)
        self.phi_p = Parameter(phi_p, device=device)
        self.phi_g = Parameter(phi_g, device=device)
        lb = lb[None, :].to(device=device)
        ub = ub[None, :].to(device=device)
        length = ub - lb
        population = torch.rand(self.pop_size, self.dim, device=device)
        population = length * population + lb
        velocity = torch.rand(self.pop_size, self.dim, device=device)
        velocity = 2 * length * velocity - length
        self.lb = lb
        self.ub = ub
        # Mutable parameters
        self.population = Mutable(population)
        self.velocity = Mutable(velocity)
        self.local_best_location = Mutable(population)
        self.local_best_fitness = Mutable(torch.empty(self.pop_size, device=device).fill_(torch.inf))
        self.global_best_location = Mutable(population[0])
        self.global_best_fitness = Mutable(torch.tensor(torch.inf, device=device))

    def step(self):
        # Compute fitness
        fitness = self.evaluate(self.population)

        # Update the local best fitness and the global best fitness
        compare = self.local_best_fitness - fitness
        self.local_best_location = torch.where(compare[:, None] > 0, self.population, self.local_best_location)
        self.local_best_fitness = self.local_best_fitness - torch.relu(compare)
        self.global_best_location, self.global_best_fitness = self._min_by(
            [self.global_best_location.unsqueeze(0), self.population],
            [self.global_best_fitness.unsqueeze(0), fitness],
        )

        # Update the velocity
        rg = torch.rand(self.pop_size, self.dim, dtype=fitness.dtype, device=fitness.device)
        rp = torch.rand(self.pop_size, self.dim, dtype=fitness.dtype, device=fitness.device)
        velocity = (
            self.w * self.velocity
            + self.phi_p * rp * (self.local_best_location - self.population)
            + self.phi_g * rg * (self.global_best_location - self.population)
        )

        # Update the population
        population = self.population + velocity
        self.population = clamp(population, self.lb, self.ub)
        self.velocity = clamp(velocity, self.lb, self.ub)

    def _min_by(
        self,
        values: List[torch.Tensor],
        keys: List[torch.Tensor],
    ):
        # Find the value with the minimum key
        values = torch.cat(values, dim=0)
        keys = torch.cat(keys, dim=0)
        min_index = torch.argmin(keys)
        return values[min_index], keys[min_index]

问题示例:Sphere 问题#

Sphere问题是一个简单但基本的基准优化问题,用于测试优化算法。

The Sphere function is defined as:

\[::\]

以下是EvoX中Sphere问题的实现示例:

class Sphere(Problem):
    def __init__(self):
        super().__init__()

    def evaluate(self, pop: torch.Tensor):
        return (pop**2).sum(-1)

使用该算法解决该问题#

初始化algorithm,problem,monitor对象。#

algorithm = PSO(
    pop_size=100,
    lb=torch.tensor([-10.0]),
    ub=torch.tensor([10.0]),
    w=0.6,
    phi_p=2.5,
    phi_g=0.8,
)
problem = Sphere()
monitor = EvalMonitor()

启动工作流并运行它#

workflow = StdWorkflow()
workflow.setup(algorithm=algorithm, problem=problem, monitor=monitor)

for _ in range(100):
    workflow.step()
workflow.get_submodule("monitor").plot()