pyduino.optimization.gradient_descent

 1import numpy as np
 2from . import Optimizer, linmap
 3
 4class GradientDescent(Optimizer):
 5    def __init__(self, population_size: int, ranges: list[float], damping: float = 0.01, rng_seed: int = 0):
 6        r"""
 7        This gradient descent algorithm assumes that the optimization function 'f' is to be minimized, differentiable, and time independent.
 8
 9        $$\frac{\mathrm{d} f}{\mathrm{d} t} = \frac{\partial f}{\partial \vec{x}}\frac{\mathrm{d} \vec{x}}{\mathrm{d} t}$$
10
11        Where $\frac{\partial f}{\partial t}$ is assumed to be zero.
12
13        Args:
14            ranges (list of pairs): Maxima and minima that each parameter can cover.
15                Example: [(0, 10), (0, 10)] for two parameters ranging from 0 to 10.
16            population_size (int): Number of individuals in the population.
17            damping (float, optional): Damping factor to avoid oscillations. Defaults to 0.01.
18            rng_seed (int, optional): Seed for the random number generator. Defaults to 0.
19        """
20        self.population_size = population_size
21        self.ranges = np.array(ranges)
22        self.damping = damping
23        self.rng_seed = np.random.default_rng(rng_seed)
24
25        # Derived attributes
26        self.invlinmap = linmap(self.ranges, np.array([[0, 1]] * len(self.ranges)))
27        self.linmap = linmap(np.array([[0, 1]] * len(self.ranges)), self.ranges)
28
29        # Initialize the population (random position and initial momentum)
30        self.population = self.rng_seed.random((self.population_size, len(self.ranges)))
31        self.momenta = self.rng_seed.random((self.population_size, len(self.ranges)))
32        self.oracle_past = self.rng_seed.random((self.population_size, 1))
33
34    def view(self, x):
35        """
36        Maps the input from the domain to the codomain.
37        """
38        return self.linmap(x)
39    
40    def view_g(self):
41        """
42        Maps the input from the domain to the codomain.
43        """
44        return self.linmap(self.population)
45
46    def inverse_view(self, x):
47        """
48        Maps the input from the codomain to the domain.
49        """
50        return self.invlinmap(x)
51    
52    def ask_oracle(self) -> np.ndarray:
53        return super().ask_oracle()
54    
55    def set_oracle(self, X: np.ndarray):
56        return super().set_oracle(X)
57
58    def init_oracle(self):
59        return self.set_oracle(self.view(self.population))
60
61    def step(self, dt: float):
62        """
63        Moves the population in the direction of the gradient.
64
65        dt float:
66            Time taken from the last observation.
67        """
68        pass