【强化学习】python 实现 saras lambda 例一

本文作者：hhh5460

本文地址：https://www.cnblogs.com/hhh5460/p/10147265.html

将例一用saras lambda算法重新撸了一遍，没有参照任何其他人的代码。仅仅根据伪代码，就撸出来了。感觉已真正理解了saras lambda算法。记录如下

0. saras lambda算法伪代码

图片来源：https://morvanzhou.github.io/static/results/reinforcement-learning/3-3-1.png（莫凡）

1. saras lambda算法真实代码

# e_table是q_table的拷贝
e_table = q_table.copy()

# ...

# saras(lambda)算法
# 参见：https://morvanzhou.github.io/static/results/reinforcement-learning/3-3-1.png
for i in range(13):
    # 0. e_table清零
    e_table *= 0
    # 1.从状态0开始
    current_state = 0
    # 2.选择一个合法的动作
    current_action = choose_action(current_state, epsilon)
    # 3.进入循环，探索学习
    while current_state != states[-1]:
        # 4.取下一状态
        next_state = get_next_state(current_state, current_action)
        # 5.取下一奖励
        next_reward = rewards[next_state]
        # 6.取下一动作
        next_action = choose_action(next_state, epsilon)
        # 7.计算德塔
        delta = next_reward + gamma * q_table.ix[next_state, next_action] - q_table.ix[current_state, current_action]
        # 8.当前状态、动作对应的e_table的值加1
        #e_table.ix[current_state, current_action] += 1 # 这是标准的操作，但是莫凡指出改成下面两句效果更好！
        e_table.ix[current_state] *= 0
        e_table.ix[current_state, current_action] = 1
        # 9.遍历每一个状态的所有动作(不能仅合法动作)
        for state in states:
            for action in actions:
                # 10.逐个更新q_talbe, e_table中对应的值
                q_table.ix[state, action] += alpha * delta * e_table.ix[state, action]
                e_table.ix[state, action] *= gamma * lambda_
        # 11.进入下一状态、动作
        current_state, current_action = next_state, next_action

第9步，刚开始我这么写：for action in get_valid_actions(state):，运行后发现没有这样写好：for action in actions:

2. 完整代码

'''
-o---T
# T 就是宝藏的位置, o 是探索者的位置
'''
# 作者: hhh5460
# 时间：20181220

'''saras(lambda)算法实现'''

import pandas as pd
import random
import time

epsilon = 0.9   # 贪婪度 greedy
alpha = 0.1     # 学习率
gamma = 0.8     # 奖励递减值
lambda_ = 0.9   # 衰减值

states = range(6)           # 状态集。从0到5
actions = ['left', 'right'] # 动作集。也可添加动作'none'，表示停留
rewards = [0,0,0,0,0,1]     # 奖励集。只有最后的宝藏所在位置才有奖励1，其他皆为0

q_table = pd.DataFrame(data=[[0 for _ in actions] for _ in states],
                       index=states, columns=actions)

e_table = q_table.copy()

def update_env(state):
    '''更新环境，并打印'''
    env = list('-----T') # 环境

    env[state] = 'o' # 更新环境
    print('\r{}'.format(''.join(env)), end='')
    time.sleep(0.1)

def get_next_state(state, action):
    '''对状态执行动作后，得到下一状态'''
    global states
    # l,r,n = -1,+1,0
    if action == 'right' and state != states[-1]: # 除末状态（位置），向右+1
        next_state = state + 1
    elif action == 'left' and state != states[0]: # 除首状态（位置），向左-1
        next_state = state -1
    else:
        next_state = state
    return next_state

def get_valid_actions(state):
    '''取当前状态下的合法动作集合，与reward无关！'''
    global actions # ['left', 'right']
    valid_actions = set(actions)
    if state == states[0]:              # 首状态（位置），则 不能向左
        valid_actions -= set(['left'])
    if state == states[-1]:             # 末状态（位置），则 不能向右
        valid_actions -= set(['right'])
    return list(valid_actions)

def choose_action(state, epsilon_=0.9):
    '''选择动作，根据状态'''
    if random.uniform(0,1) > epsilon_: # 探索
        action = random.choice(get_valid_actions(state))
    else:                             # 利用（贪婪）
        #current_action = q_table.ix[current_state].idxmax() # 这种写法是有问题的！
        s = q_table.ix[state].filter(items=get_valid_actions(state))
        action = random.choice(s[s==s.max()].index) # 可能多个最大值，当然，一个更好
    return action

# saras(lambda)算法
# 参见：https://morvanzhou.github.io/static/results/reinforcement-learning/3-3-1.png
for i in range(13):
    e_table *= 0 # 清零

    current_state = 0
    current_action = choose_action(current_state, epsilon)

    update_env(current_state) # 环境相关
    total_steps = 0           # 环境相关

    while current_state != states[-1]:
        next_state = get_next_state(current_state, current_action)
        next_reward = rewards[next_state]

        next_action = choose_action(next_state, epsilon)
        delta = next_reward + gamma * q_table.ix[next_state, next_action] - q_table.ix[current_state, current_action]
        #e_table.ix[current_state, current_action] += 1 # 这是标准的操作，但是莫凡指出改成下面两句效果更好！
        e_table.ix[current_state] *= 0
        e_table.ix[current_state, current_action] = 1
        for state in states:
            for action in actions: #get_valid_actions(state):
                q_table.ix[state, action] += alpha * delta * e_table.ix[state, action]
                e_table.ix[state, action] *= gamma * lambda_
        current_state, current_action = next_state, next_action

        update_env(current_state) # 环境相关
        total_steps += 1          # 环境相关

    print('\rEpisode {}: total_steps = {}'.format(i, total_steps), end='') # 环境相关
    time.sleep(2)                                                          # 环境相关
    print('\r                                ', end='')                    # 环境相关

print('\nq_table:')
print(q_table)