agent.py

from math import sin, cos, pi, exp
import numpy as np
import ImageDraw


def getDDvf(im, x, y, theta, epsilon):
    '''Approximate (w/ epsilon) the directional derivative of f at (x,y)
    along a vector defined by theta.'''
    x2 = x + epsilon * cos(theta)
    y2 = y + epsilon * sin(theta)

    x = max(x, 0)
    y = max(y, 0)

    x2 = max(x2, 0)
    y2 = max(y2, 0)

    x2 = min(x2, im.shape[1] - 1)
    y2 = min(y2, im.shape[0] - 1)

    x2 = round(x2)
    y2 = round(y2)

    p1 = im[y, x]
    p2 = im[y2, x2]
    return (p2 - p1) / epsilon


def getX(im, x, y, thetas, epsilon):
    '''Get gradient for n evenly-spaced radial directions around [x,y]'''
    return [getDDvf(im, x, y, theta, epsilon) for theta in thetas]


def h_ax(alpha, x):
    '''h_alpha(x) = 2*pi(sigmoid(x)-0.5)'''
    return 2 * pi * (1.0 / (1 + exp(np.dot(alpha, x))) - 0.5)


class Agent:
    def __init__(self, alpha):
        self.alpha = alpha
        self.paths = None
        self.finished = None

    def render_paths(self, im):
        path_img = im.copy()
        draw = ImageDraw.Draw(path_img)
        for path in self.paths:
            draw.line(path, fill=(255, 0, 0))
        return path_img

    def reset_paths(self):
        self.paths = []

    def score(self, im, x, y, epsilon, lam, n, maxN, Q):
        '''Compute the score of this agent on the provided image.'''
        thetas = np.linspace(0, 2 * np.pi, n+1)[:-1]
        path = [(x, y)]
        cost = 0.0
        self.finished = False
        for i in range(maxN):
            # Compute model output
            X = getX(im, x, y, thetas, epsilon)
            phi_hat = h_ax(self.alpha, X)

            # Compute next position based on model output
            x_next = round(x + epsilon * cos(phi_hat))
            y_next = round(y + epsilon * sin(phi_hat))

            # Clamp next position to the image boundary
            x_next = np.clip(x_next, 0, im.shape[1] - 1)
            y_next = np.clip(y_next, 0, im.shape[0] - 1)

            # Update the cost
            a = im[y, x]
            b = im[y_next, x_next]
            delta_n = float(b - a)
            if delta_n >= 0:
                g_n = delta_n
            else:
                g_n = - delta_n / 2
            cost += (g_n ** 2)

            path.append((x_next, y_next))

            # Stop if we've reached the goal
            if x_next == im.shape[1] - 1:
                self.finished = True
                break

            x, y = x_next, y_next

        # Add the penalty if the goal wasn't reached
        if not self.finished:
            cost += Q

        cost += lam * len(path)

        self.paths.append(path)

        return cost