核心代码
def sample(self, obs):
# 根据输入观察值 生成一个动作输出 带探索
greed = np.random.uniform(0, 1) < (1 - self.epsilon)
if greed:
action = self.predict(obs)
else:
action = np.random.choice(self.act_n)
return action
def predict(self, obs):
# 获取表中列表
obs_lst = self.Q[obs, :] # 状态列表
# 从列表中选择动作最大值(可能有几个一样大)
obs_max = np.max(obs_lst)
max_action_lst = np.where(obs_lst == obs_max)[0] # 获取最大索引(动作)
action = np.random.choice(max_action_lst) # 随机选择 动作
return action
# 学习 更新Q表格
def learn(self, obs, action, reward, next_obs, next_action, done):
""" on-policy
obs: 交互前的obs, s_t
action: 本次交互选择的action, a_t
reward: 本次动作获得的奖励r
next_obs: 本次交互后的obs, s_t+1
next_action: 根据当前Q表格, 针对next_obs会选择的动作, a_t+1
done: episode是否结束
"""
# 获取当前值
obs_act_q = self.Q[obs, action]
# 计算 下次目标值
if done:
target_q = reward
else:
target_q = reward + self.gamma * self.Q[next_obs, next_action]
# sarsa s:当前状态 a 当前动作 r: 反馈 s:下个一个状态 a:下一个动作
# 更新当前值
self.Q[obs, action] += self.lr * (target_q - obs_act_q)
完整代码
主程序
import gym
from gridworld import CliffWalkingWapper
from sarsa_agent import SarsaAgent
import time
# 运行
def run_episode(env, agent, render=False):
total_steps = 0
total_reward = 0
obs = env.reset()
action = agent.sample(obs) # 选择一个动作
while True:
# 获取 s = obs ; a = action r = reward s = next_obs; a = next_action
next_obs, reward, done, _ = env.step(action)
# 有一定几率 随机动作
next_action = agent.sample(next_obs) # 0 up, 1 right, 2 down, 3 left
# print('学习数据', obs, action, reward, next_obs, next_action, done)
agent.learn(obs, action, reward, next_obs, next_action, done)
action = next_action
obs = next_obs
total_reward += reward
total_steps += 1
if render:
env.render()
if done:
break
return total_reward, total_steps
def test_episode(env, agent):
total_reward = 0
obs = env.reset()
while True:
action = agent.predict(obs) # greedy
next_obs, reward, done, _ = env.step(action)
total_reward += reward
obs = next_obs
time.sleep(0.5)
env.render()
if done:
break
return total_reward
if __name__ == '__main__':
env = gym.make('CliffWalking-v0')
env = CliffWalkingWapper(env)
# 创建一个agent实例,输入超参数
agent = SarsaAgent(
obs_n=env.observation_space.n,
act_n=env.action_space.n,
learning_rate=0.1,
gamma=0.9,
e_greed=0.1)
# 训练500个episode,打印每个20 episode的分数
for episode in range(500):
render = True if episode % 20 == 0 else False
ep_reward, ep_steps = run_episode(env, agent, render)
print('Episode %s: steps = %s , reward = %.1f' % (episode, ep_steps, ep_reward))
# 全部训练结束,查看算法效果
test_reward = test_episode(env, agent)
print('test reward = %.1f' % test_reward)
agent.save()
sarsa_agent.py
import numpy as np
class SarsaAgent(object):
def __init__(self, obs_n, act_n, learning_rate=0.01, gamma=0.9, e_greed=0.1):
self.act_n = act_n # 动作的可能性
self.obs_n = obs_n # 环境状态数
self.lr = learning_rate # 学习率
self.gamma = gamma # 衰减率
self.epsilon = e_greed # 概率随机动作
self.Q = np.zeros((obs_n, act_n)) # 生成Q表格
def sample(self, obs):
# 根据输入观察值 生成一个动作输出 带探索
greed = np.random.uniform(0, 1) < (1 - self.epsilon)
if greed:
action = self.predict(obs)
else:
action = np.random.choice(self.act_n)
return action
# 预测输出的值
def predict(self, obs):
# 获取表中列表
obs_lst = self.Q[obs, :] # 状态列表
# 从列表中选择动作最大值(可能有几个一样大)
obs_max = np.max(obs_lst)
max_action_lst = np.where(obs_lst == obs_max)[0] # 获取最大索引(动作)
action = np.random.choice(max_action_lst) # 随机选择 动作
return action
# 学习 更新Q表格
def learn(self, obs, action, reward, next_obs, next_action, done):
""" on-policy
obs: 交互前的obs, s_t
action: 本次交互选择的action, a_t
reward: 本次动作获得的奖励r
next_obs: 本次交互后的obs, s_t+1
next_action: 根据当前Q表格, 针对next_obs会选择的动作, a_t+1
done: episode是否结束
"""
# 获取当前值
obs_act_q = self.Q[obs, action]
# 计算 下次目标值
if done:
target_q = reward
else:
target_q = reward + self.gamma * self.Q[next_obs, next_action]
# sarsa s:当前状态 a 当前动作 r: 反馈 s:下个一个状态 a:下一个动作
# 更新当前值
self.Q[obs, action] += self.lr * (target_q - obs_act_q)
def save(self):
npy_file = './q_table.npy'
np.save(npy_file, self.Q)
print(npy_file + ' saved.')
def restore(self, npy_file='./q_table.npy'):
self.Q = np.load(npy_file)
print(npy_file + ' loaded.')
gridworld.py
# -*- coding: utf-8 -*-
import gym
import turtle
import numpy as np
# turtle tutorial : https://docs.python.org/3.3/library/turtle.html
def GridWorld(gridmap=None, is_slippery=False):
if gridmap is None:
gridmap = ['SFFF', 'FHFH', 'FFFH', 'HFFG']
env = gym.make("FrozenLake-v0", desc=gridmap, is_slippery=False)
env = FrozenLakeWapper(env)
return env
class FrozenLakeWapper(gym.Wrapper):
def __init__(self, env):
gym.Wrapper.__init__(self, env)
self.max_y = env.desc.shape[0]
self.max_x = env.desc.shape[1]
self.t = None
self.unit = 50
def draw_box(self, x, y, fillcolor='', line_color='gray'):
self.t.up()
self.t.goto(x * self.unit, y * self.unit)
self.t.color(line_color)
self.t.fillcolor(fillcolor)
self.t.setheading(90)
self.t.down()
self.t.begin_fill()
for _ in range(4):
self.t.forward(self.unit)
self.t.right(90)
self.t.end_fill()
def move_player(self, x, y):
self.t.up()
self.t.setheading(90)
self.t.fillcolor('red')
self.t.goto((x + 0.5) * self.unit, (y + 0.5) * self.unit)
def render(self):
if self.t == None:
self.t = turtle.Turtle()
self.wn = turtle.Screen()
self.wn.setup(self.unit * self.max_x + 100,
self.unit * self.max_y + 100)
self.wn.setworldcoordinates(0, 0, self.unit * self.max_x,
self.unit * self.max_y)
self.t.shape('circle')
self.t.width(2)
self.t.speed(0)
self.t.color('gray')
for i in range(self.desc.shape[0]):
for j in range(self.desc.shape[1]):
x = j
y = self.max_y - 1 - i
if self.desc[i][j] == b'S': # Start
self.draw_box(x, y, 'white')
elif self.desc[i][j] == b'F': # Frozen ice
self.draw_box(x, y, 'white')
elif self.desc[i][j] == b'G': # Goal
self.draw_box(x, y, 'yellow')
elif self.desc[i][j] == b'H': # Hole
self.draw_box(x, y, 'black')
else:
self.draw_box(x, y, 'white')
self.t.shape('turtle')
x_pos = self.s % self.max_x
y_pos = self.max_y - 1 - int(self.s / self.max_x)
self.move_player(x_pos, y_pos)
class CliffWalkingWapper(gym.Wrapper):
def __init__(self, env):
gym.Wrapper.__init__(self, env)
self.t = None
self.unit = 50
self.max_x = 12
self.max_y = 4
def draw_x_line(self, y, x0, x1, color='gray'):
assert x1 > x0
self.t.color(color)
self.t.setheading(0)
self.t.up()
self.t.goto(x0, y)
self.t.down()
self.t.forward(x1 - x0)
def draw_y_line(self, x, y0, y1, color='gray'):
assert y1 > y0
self.t.color(color)
self.t.setheading(90)
self.t.up()
self.t.goto(x, y0)
self.t.down()
self.t.forward(y1 - y0)
def draw_box(self, x, y, fillcolor='', line_color='gray'):
self.t.up()
self.t.goto(x * self.unit, y * self.unit)
self.t.color(line_color)
self.t.fillcolor(fillcolor)
self.t.setheading(90)
self.t.down()
self.t.begin_fill()
for i in range(4):
self.t.forward(self.unit)
self.t.right(90)
self.t.end_fill()
def move_player(self, x, y):
self.t.up()
self.t.setheading(90)
self.t.fillcolor('red')
self.t.goto((x + 0.5) * self.unit, (y + 0.5) * self.unit)
def render(self):
if self.t == None:
self.t = turtle.Turtle()
self.wn = turtle.Screen()
self.wn.setup(self.unit * self.max_x + 100,
self.unit * self.max_y + 100)
self.wn.setworldcoordinates(0, 0, self.unit * self.max_x,
self.unit * self.max_y)
self.t.shape('circle')
self.t.width(2)
self.t.speed(0)
self.t.color('gray')
for _ in range(2):
self.t.forward(self.max_x * self.unit)
self.t.left(90)
self.t.forward(self.max_y * self.unit)
self.t.left(90)
for i in range(1, self.max_y):
self.draw_x_line(
y=i * self.unit, x0=0, x1=self.max_x * self.unit)
for i in range(1, self.max_x):
self.draw_y_line(
x=i * self.unit, y0=0, y1=self.max_y * self.unit)
for i in range(1, self.max_x - 1):
self.draw_box(i, 0, 'black')
self.draw_box(self.max_x - 1, 0, 'yellow')
self.t.shape('turtle')
x_pos = self.s % self.max_x
y_pos = self.max_y - 1 - int(self.s / self.max_x)
self.move_player(x_pos, y_pos)
if __name__ == '__main__':
# 环境1:FrozenLake, 可以配置冰面是否是滑的
# 0 left, 1 down, 2 right, 3 up
env = gym.make("FrozenLake-v0", is_slippery=False)
env = FrozenLakeWapper(env)
# 环境2:CliffWalking, 悬崖环境
# env = gym.make("CliffWalking-v0") # 0 up, 1 right, 2 down, 3 left
# env = CliffWalkingWapper(env)
# 环境3:自定义格子世界,可以配置地图, S为出发点Start, F为平地Floor, H为洞Hole, G为出口目标Goal
# gridmap = [
# 'SFFF',
# 'FHFF',
# 'FFFF',
# 'HFGF' ]
# env = GridWorld(gridmap)
env.reset()
for step in range(10):
action = np.random.randint(0, 4)
obs, reward, done, info = env.step(action)
print('step {}: action {}, obs {}, reward {}, done {}, info {}'.format(\
step, action, obs, reward, done, info))
# env.render() # 渲染一帧图像