Sarsa（lambda）理解

上代码

import numpy as np
import pandas as pd


class RL(object):
    def __init__(self, action_space, learning_rate=0.01, reward_decay=0.9, e_greedy=0.9):
        self.actions = action_space  # a list
        self.lr = learning_rate
        self.gamma = reward_decay
        self.epsilon = e_greedy

        self.q_table = pd.DataFrame(columns=self.actions, dtype=np.float64)

    def check_state_exist(self, state):
        if state not in self.q_table.index:
            # append new state to q table
            self.q_table = self.q_table.append(
                pd.Series(
                    [0]*len(self.actions),
                    index=self.q_table.columns,
                    name=state,
                )
            )

    def choose_action(self, observation):
        self.check_state_exist(observation)
        # action selection
        if np.random.rand() < self.epsilon:
            # choose best action
            state_action = self.q_table.loc[observation, :]
            # some actions may have the same value, randomly choose on in these actions
            action = np.random.choice(state_action[state_action == np.max(state_action)].index)
        else:
            # choose random action
            action = np.random.choice(self.actions)
        return action

    def learn(self, *args):
        pass


# backward eligibility traces
class SarsaLambdaTable(RL):
    def __init__(self, actions, learning_rate=0.01, reward_decay=0.9, e_greedy=0.9, trace_decay=0.9):
        super(SarsaLambdaTable, self).__init__(actions, learning_rate, reward_decay, e_greedy)

        # backward view, eligibility trace.
        self.lambda_ = trace_decay
        self.eligibility_trace = self.q_table.copy()

    def check_state_exist(self, state):
        if state not in self.q_table.index:
            # append new state to q table
            to_be_append = pd.Series(
                    [0] * len(self.actions),
                    index=self.q_table.columns,
                    name=state,
                )
            self.q_table = self.q_table.append(to_be_append)

            # also update eligibility trace
            self.eligibility_trace = self.eligibility_trace.append(to_be_append)

    def learn(self, s, a, r, s_, a_):
        self.check_state_exist(s_)
        q_predict = self.q_table.loc[s, a]
        if s_ != 'terminal':
            q_target = r + self.gamma * self.q_table.loc[s_, a_]  # next state is not terminal
        else:
            q_target = r  # next state is terminal
        error = q_target - q_predict

        # increase trace amount for visited state-action pair

        # Method 1:
        # self.eligibility_trace.loc[s, a] += 1

        # Method 2:
        self.eligibility_trace.loc[s, :] *= 0
        self.eligibility_trace.loc[s, a] = 1

        # Q update
        self.q_table += self.lr * error * self.eligibility_trace

        # decay eligibility trace after update
        self.eligibility_trace *= self.gamma*self.lambda_

--------------------------------------------------------------------

from maze_env import Maze
from RL_brain import SarsaLambdaTable


def update():
    for episode in range(100):
        # initial observation
        observation = env.reset()

        # RL choose action based on observation
        action = RL.choose_action(str(observation))

        # initial all zero eligibility trace
        RL.eligibility_trace *= 0

        while True:
            # fresh env
            env.render()

            # RL take action and get next observation and reward
            observation_, reward, done = env.step(action)

            # RL choose action based on next observation
            action_ = RL.choose_action(str(observation_))

            # RL learn from this transition (s, a, r, s, a) ==> Sarsa
            RL.learn(str(observation), action, reward, str(observation_), action_)

            # swap observation and action
            observation = observation_
            action = action_

            # break while loop when end of this episode
            if done:
                break

    # end of game
    print('game over')
    env.destroy()

if __name__ == "__main__":
    env = Maze()
    RL = SarsaLambdaTable(actions=list(range(env.n_actions)))

    env.after(100, update)
    env.mainloop()

 

• 在每次take action获得reward后，Sarsa只对前一步Q(s,a)进行更新，Sarsa(lambda) 则会对获得reward之前的步进行更新。

arsa 是一种单步更新法, 也就是 Sarsa(0), 因为他等走完这一步以后直接更新行为准则. 如果延续这种想法, 走完这步, 再走一步, 然后再更新, 我们可以叫他 Sarsa(1). 同理, 如果等待回合完毕我们一次性再更新呢, 比如这回合我们走了 n 步, 那我们就叫 Sarsa(n). 为了统一这样的流程, 我们就有了一个 lambda 值来代替我们想要选择的步数。 这也就是 Sarsa(λ)的由来，Sarsa 和 Qlearning 都是每次获取到奖励reward后只更新获取到 reward 的前一步，那么Sarsa(λ)就是更新获取到 reward 的前 λ 步. λ 在 [0, 1] 之间取值, 
当 lambda = 0, Sarsa-lambda 就是 Sarsa单步更新, 只更新获取到 reward 前经历的最后一步。如果 lambda = 1, Sarsa-lambda就变成了回合更新，更新的是获取到 reward 前所有经历的步，对所有步更新的力度都是一样. 当 lambda 在 0 和 1 之间, 取值越大, 获得奖励大的步更新力度越大. 这样我们就不用受限于单步更新的每次只能更新最近的一步, 我们可以更有效率的更新所有相关步了。

 

第一种  路过就在矩阵中+1   ，第二种持续加但是衰变慢，并且路过的次数越多就代表越重要

第三种  限制最大值，减少重复步骤的作用

 

 这就是脚步之所以会衰变的原因，因为初始E一直会衰变，只有有很大作用的值会留下来

两种E更新方式。（1）路过就加一，无上限，数值越来证明在此状态下执行这个动作的次数越大

（2）始终唯一，只记录走过的路，而不管哪一步被走了多少次。