强化学习Q-Learning算法详解

3605 人阅读 | 时间：2019年10月29日 10:30

强化学习Q-Learning算法详解

https://study.163.com/provider/400000000398149/index.htm?share=2&shareId=400000000398149（欢迎关注博主主页，学习python视频资源，还有大量免费python经典文章）

强化学习Q-Learning算法详解

【强化学习】Q-Learning详解
1、算法思想
QLearning是强化学习算法中值迭代的算法，Q即为Q（s,a）就是在某一时刻的s 状态下(s∈S)，采取a (a∈A)动作能够获得收益的期望，环境会根据agent的动作反馈相应的回报reward r，所以算法的主要思想就是将State与Action构建成一张Q-table来存储Q值，然后根据Q值来选取动作获得较大的收益。

2、公式推导
举个例子如图有一个GridWorld的游戏从起点出发到达终点为胜利掉进陷阱为失败。智能体（Agent）、环境状态（environment）、奖励（reward）、动作（action）可以将问题抽象成一个马尔科夫决策过程，我们在每个格子都算是一个状态$s_t $ , π(a|s)在s状态下采取动作aa∈A 。P(s’|s,a)为在s状态下选择a动作转换到下一个状态s’的概率。R(s’|s,a)表示在s状态下采取a动作转移到s’的奖励reward，我们的目的很明确就是找到一条能够到达终点获得最大奖赏的策略。

所以目标就是求出累计奖赏最大的策略的期望：

强化学习Q-Learning算法详解

4、实现代码

值迭代部分

# -*- coding: utf-8 -*-
from environment import GraphicDisplay, Env
class ValueIteration:
    def __init__(self, env):
        self.env = env
        # 2-d list for the value function
        self.value_table = [[0.0] * env.width for _ in range(env.height)]
        self.discount_factor = 0.9
    # get next value function table from the current value function table
    def value_iteration(self):
        next_value_table = [[0.0] * self.env.width
                                    for _ in range(self.env.height)]
        for state in self.env.get_all_states():
            if state == [2, 2]:
                next_value_table[state[0]][state[1]] = 0.0
                continue
            value_list = []
            for action in self.env.possible_actions:
                next_state = self.env.state_after_action(state, action)
                reward = self.env.get_reward(state, action)
                next_value = self.get_value(next_state)
                value_list.append((reward + self.discount_factor * next_value))
            # return the maximum value(it is the optimality equation!!)
            next_value_table[state[0]][state[1]] = round(max(value_list), 2)
        self.value_table = next_value_table
    # get action according to the current value function table
    def get_action(self, state):
        action_list = []
        max_value = -99999
        if state == [2, 2]:
            return []
        # calculating q values for the all actions and
        # append the action to action list which has maximum q value
        for action in self.env.possible_actions:
            next_state = self.env.state_after_action(state, action)
            reward = self.env.get_reward(state, action)
            next_value = self.get_value(next_state)
            value = (reward + self.discount_factor * next_value)
            if value > max_value:
                action_list.clear()
                action_list.append(action)
                max_value = value
            elif value == max_value:
                action_list.append(action)
        return action_list
    def get_value(self, state):
        return round(self.value_table[state[0]][state[1]], 2)
if __name__ == "__main__":
    env = Env()
    value_iteration = ValueIteration(env)
    grid_world = GraphicDisplay(value_iteration)
    grid_world.mainloop()

动态环境部分

import tkinter as tk
import time
import numpy as np
import random
from PIL import ImageTk, Image
PhotoImage = ImageTk.PhotoImage
UNIT = 100  # pixels
HEIGHT = 5  # grid height
WIDTH = 5  # grid width
TRANSITION_PROB = 1
POSSIBLE_ACTIONS = [0, 1, 2, 3]  # up, down, left, right
ACTIONS = [(-1, 0), (1, 0), (0, -1), (0, 1)]  # actions in coordinates
REWARDS = []
class GraphicDisplay(tk.Tk):
    def __init__(self, value_iteration):
        super(GraphicDisplay, self).__init__()
        self.title('Value Iteration')
        self.geometry('{0}x{1}'.format(HEIGHT * UNIT, HEIGHT * UNIT + 50))
        self.texts = []
        self.arrows = []
        self.env = Env()
        self.agent = value_iteration
        self.iteration_count = 0
        self.improvement_count = 0
        self.is_moving = 0
        (self.up, self.down, self.left,
         self.right), self.shapes = self.load_images()
        self.canvas = self._build_canvas()
        self.text_reward(2, 2, "R : 1.0")
        self.text_reward(1, 2, "R : -1.0")
        self.text_reward(2, 1, "R : -1.0")
    def _build_canvas(self):
        canvas = tk.Canvas(self, bg='white',
                           height=HEIGHT * UNIT,
                           width=WIDTH * UNIT)
        # buttons
        iteration_button = tk.Button(self, text="Calculate",
                                     command=self.calculate_value)
        iteration_button.configure(width=10, activebackground="#33B5E5")
        canvas.create_window(WIDTH * UNIT * 0.13, (HEIGHT * UNIT) + 10,
                             window=iteration_button)
        policy_button = tk.Button(self, text="Print Policy",
                                  command=self.print_optimal_policy)
        policy_button.configure(width=10, activebackground="#33B5E5")
        canvas.create_window(WIDTH * UNIT * 0.37, (HEIGHT * UNIT) + 10,
                             window=policy_button)
        policy_button = tk.Button(self, text="Move",
                                  command=self.move_by_policy)
        policy_button.configure(width=10, activebackground="#33B5E5")
        canvas.create_window(WIDTH * UNIT * 0.62, (HEIGHT * UNIT) + 10,
                             window=policy_button)
        policy_button = tk.Button(self, text="Clear", command=self.clear)
        policy_button.configure(width=10, activebackground="#33B5E5")
        canvas.create_window(WIDTH * UNIT * 0.87, (HEIGHT * UNIT) + 10,
                             window=policy_button)
        # create grids
        for col in range(0, WIDTH * UNIT, UNIT):  # 0~400 by 80
            x0, y0, x1, y1 = col, 0, col, HEIGHT * UNIT
            canvas.create_line(x0, y0, x1, y1)
        for row in range(0, HEIGHT * UNIT, UNIT):  # 0~400 by 80
            x0, y0, x1, y1 = 0, row, HEIGHT * UNIT, row
            canvas.create_line(x0, y0, x1, y1)
        # add img to canvas
        self.rectangle = canvas.create_image(50, 50, image=self.shapes[0])
        canvas.create_image(250, 150, image=self.shapes[1])
        canvas.create_image(150, 250, image=self.shapes[1])
        canvas.create_image(250, 250, image=self.shapes[2])
        # pack all
        canvas.pack()
        return canvas
    def load_images(self):
        PhotoImage = ImageTk.PhotoImage
        up = PhotoImage(Image.open("../img/up.png").resize((13, 13)))
        right = PhotoImage(Image.open("../img/right.png").resize((13, 13)))
        left = PhotoImage(Image.open("../img/left.png").resize((13, 13)))
        down = PhotoImage(Image.open("../img/down.png").resize((13, 13)))
        rectangle = PhotoImage(
            Image.open("../img/rectangle.png").resize((65, 65)))
        triangle = PhotoImage(
            Image.open("../img/triangle.png").resize((65, 65)))
        circle = PhotoImage(Image.open("../img/circle.png").resize((65, 65)))
        return (up, down, left, right), (rectangle, triangle, circle)
    def clear(self):
        if self.is_moving == 0:
            self.iteration_count = 0
            self.improvement_count = 0
            for i in self.texts:
                self.canvas.delete(i)
            for i in self.arrows:
                self.canvas.delete(i)
            self.agent.value_table = [[0.0] * WIDTH for _ in range(HEIGHT)]
            x, y = self.canvas.coords(self.rectangle)
            self.canvas.move(self.rectangle, UNIT / 2 - x, UNIT / 2 - y)
    def reset(self):
        self.update()
        time.sleep(0.5)
        self.canvas.delete(self.rectangle)
        return self.canvas.coords(self.rectangle)
    def text_value(self, row, col, contents, font='Helvetica', size=12,
                   style='normal', anchor="nw"):
        origin_x, origin_y = 85, 70
        x, y = origin_y + (UNIT * col), origin_x + (UNIT * row)
        font = (font, str(size), style)
        text = self.canvas.create_text(x, y, fill="black", text=contents,
                                       font=font, anchor=anchor)
        return self.texts.append(text)
    def text_reward(self, row, col, contents, font='Helvetica', size=12,
                    style='normal', anchor="nw"):
        origin_x, origin_y = 5, 5
        x, y = origin_y + (UNIT * col), origin_x + (UNIT * row)
        font = (font, str(size), style)
        text = self.canvas.create_text(x, y, fill="black", text=contents,
                                       font=font, anchor=anchor)
        return self.texts.append(text)
    def rectangle_move(self, action):
        base_action = np.array([0, 0])
        location = self.find_rectangle()
        self.render()
        if action == 0 and location[0] > 0:  # up
            base_action[1] -= UNIT
        elif action == 1 and location[0] < HEIGHT - 1:  # down
            base_action[1] += UNIT
        elif action == 2 and location[1] > 0:  # left
            base_action[0] -= UNIT
        elif action == 3 and location[1] < WIDTH - 1:  # right
            base_action[0] += UNIT
        self.canvas.move(self.rectangle, base_action[0],
                         base_action[1])  # move agent
    def find_rectangle(self):
        temp = self.canvas.coords(self.rectangle)
        x = (temp[0] / 100) - 0.5
        y = (temp[1] / 100) - 0.5
        return int(y), int(x)
    def move_by_policy(self):
        if self.improvement_count != 0 and self.is_moving != 1:
            self.is_moving = 1
            x, y = self.canvas.coords(self.rectangle)
            self.canvas.move(self.rectangle, UNIT / 2 - x, UNIT / 2 - y)
            x, y = self.find_rectangle()
            while len(self.agent.get_action([x, y])) != 0:
                action = random.sample(self.agent.get_action([x, y]), 1)[0]
                self.after(100, self.rectangle_move(action))
                x, y = self.find_rectangle()
            self.is_moving = 0
    def draw_one_arrow(self, col, row, action):
        if col == 2 and row == 2:
            return
        if action == 0:  # up
            origin_x, origin_y = 50 + (UNIT * row), 10 + (UNIT * col)
            self.arrows.append(self.canvas.create_image(origin_x, origin_y,
                                                        image=self.up))
        elif action == 1:  # down
            origin_x, origin_y = 50 + (UNIT * row), 90 + (UNIT * col)
            self.arrows.append(self.canvas.create_image(origin_x, origin_y,
                                                        image=self.down))
        elif action == 3:  # right
            origin_x, origin_y = 90 + (UNIT * row), 50 + (UNIT * col)
            self.arrows.append(self.canvas.create_image(origin_x, origin_y,
                                                        image=self.right))
        elif action == 2:  # left
            origin_x, origin_y = 10 + (UNIT * row), 50 + (UNIT * col)
            self.arrows.append(self.canvas.create_image(origin_x, origin_y,
                                                        image=self.left))
    def draw_from_values(self, state, action_list):
        i = state[0]
        j = state[1]
        for action in action_list:
            self.draw_one_arrow(i, j, action)
    def print_values(self, values):
        for i in range(WIDTH):
            for j in range(HEIGHT):
                self.text_value(i, j, values[i][j])
    def render(self):
        time.sleep(0.1)
        self.canvas.tag_raise(self.rectangle)
        self.update()
    def calculate_value(self):
        self.iteration_count += 1
        for i in self.texts:
            self.canvas.delete(i)
        self.agent.value_iteration()
        self.print_values(self.agent.value_table)
    def print_optimal_policy(self):
        self.improvement_count += 1
        for i in self.arrows:
            self.canvas.delete(i)
        for state in self.env.get_all_states():
            action = self.agent.get_action(state)
            self.draw_from_values(state, action)
class Env:
    def __init__(self):
        self.transition_probability = TRANSITION_PROB
        self.width = WIDTH  # Width of Grid World
        self.height = HEIGHT  # Height of GridWorld
        self.reward = [[0] * WIDTH for _ in range(HEIGHT)]
        self.possible_actions = POSSIBLE_ACTIONS
        self.reward[2][2] = 1  # reward 1 for circle
        self.reward[1][2] = -1  # reward -1 for triangle
        self.reward[2][1] = -1  # reward -1 for triangle
        self.all_state = []
        for x in range(WIDTH):
            for y in range(HEIGHT):
                state = [x, y]
                self.all_state.append(state)
    def get_reward(self, state, action):
        next_state = self.state_after_action(state, action)
        return self.reward[next_state[0]][next_state[1]]
    def state_after_action(self, state, action_index):
        action = ACTIONS[action_index]
        return self.check_boundary([state[0] + action[0], state[1] + action[1]])
    @staticmethod
    def check_boundary(state):
        state[0] = (0 if state[0] < 0 else WIDTH - 1
        if state[0] > WIDTH - 1 else state[0])
        state[1] = (0 if state[1] < 0 else HEIGHT - 1
        if state[1] > HEIGHT - 1 else state[1])
        return state
    def get_transition_prob(self, state, action):
        return self.transition_probability
    def get_all_states(self):
        return self.all_state

转载https://blog.csdn.net/qq_30615903/article/details/80739243