Code Experiment

测试环境叫做cliff run，如下图：

起始位置是左下角(3,0)，目标位置是右下角(3,11)，其中黄色方框是当前位置，深紫色的是悬崖，如果掉下去了这个回合就结束了，得重新开始。

具体的文件可以在我的Github上，下载下来，直接运行jupyter notebook即可
https://github.com/Qxxxx/ReinforcementLearning.git

Q-Learning

class QLearningAgent(Agent):
   
    def __init__(self, actions, epsilon=0.01, alpha=0.5, gamma=1):
        super(QLearningAgent, self).__init__(actions)
     
        ## Initialize empty dictionary here
        ## In addition, initialize the value of epsilon, alpha and gamma
        self.Q = {}
        self.epsilon = epsilon
        self.alpha = alpha
        self.gamma = gamma
    def stateToString(self, state):
        mystring = ""
        if np.isscalar(state):
            mystring = str(state)
        else:
            for digit in state:
                mystring += str(digit)
        return mystring    
    
    def act(self, state):
        stateStr = self.stateToString(state)      
        action = np.random.randint(0, self.num_actions) 
        
        
        Q = self.num_actions*[0]
        for a in range(self.num_actions):
            if not stateStr+' %i'%a in self.Q:
                self.Q[stateStr+' %i'%a] = 0
            Q[a] = self.Q[stateStr+' %i'%a]
        
        
        choice = None
        if self.epsilon == 0:
            choice = 0
        elif self.epsilon == 1:
            choice = 1
        else:
            choice = np.random.binomial(1, self.epsilon)
            
        if choice == 1:
            return np.random.randint(0, self.num_actions)
        else:
            m = max(Q)
            best_Q = [i for i, j in enumerate(Q) if j == m]
            action = np.random.choice(best_Q)
        #set_trace()
        return action
        return action
    
    def learn(self, state1, action1, reward, state2, done):
        state1Str = self.stateToString(state1)
        state2Str = self.stateToString(state2)
        
        
        Q = self.num_actions*[0]
        for a in range(self.num_actions):
            if not state2Str+' %i'%a in self.Q:
                self.Q[state2Str+' %i'%a] = 0
            Q[a] = self.Q[state2Str+' %i'%a]
        self.Q[state1Str+' %i'%action1] = self.Q[state1Str+' %i'%action1]+\
        self.alpha*(reward+self.gamma*max(Q)-\
                    self.Q[state1Str+' %i'%action1])
        """
        Q-learning Update:
        Q(s,a) <- Q(s,a) + alpha * (reward + gamma * max(Q(s') - Q(s,a))
        """

SARSA

class SarsaAgent(Agent):
   
    def __init__(self, actions, epsilon=0.01, alpha=0.5, gamma=1):
        super(SarsaAgent, self).__init__(actions)
        
        
        ## Initialize empty dictionary here
        ## In addition, initialize the value of epsilon, alpha and gamma
        self.Q = {}
        self.epsilon = epsilon
        self.alpha = alpha
        self.gamma = gamma
        
    def stateToString(self, state):
        mystring = ""
        if np.isscalar(state):
            mystring = str(state)
        else:
            for digit in state:
                mystring += str(digit)
        return mystring    
    
    def act(self, state):
        stateStr = self.stateToString(state)      
        action = np.random.randint(0, self.num_actions)
        Q = self.num_actions*[0]
        for a in range(self.num_actions):
            if not stateStr+' %i'%a in self.Q:
                self.Q[stateStr+' %i'%a] = 0
            Q[a] = self.Q[stateStr+' %i'%a]
        
        
        ## Implement epsilon greedy policy here
        choice = None
        if self.epsilon == 0:
            choice = 0
        elif self.epsilon == 1:
            choice = 1
        else:
            choice = np.random.binomial(1, self.epsilon)
            
        if choice == 1:
            return np.random.randint(0, self.num_actions)
        else:
            m = max(Q)
            best_Q = [i for i, j in enumerate(Q) if j == m]
            action = np.random.choice(best_Q)
        #set_trace()
        return action

    def learn(self, state1, action1, reward, state2, action2):
        state1Str = self.stateToString(state1)
        state2Str = self.stateToString(state2)
        
        
        ## Implement the sarsa update here
 
        #if not state2Str+' %i'%action2 in self.Q:
        #    self.Q[state2Str+' %i'%action2] = 0
        self.Q[state1Str+' %i'%action1] = self.Q[state1Str+' %i'%action1]+\
        self.alpha*(reward+self.gamma*self.Q[state2Str+' %i'%action2]-self.Q[state1Str+' %i'%action1])
        #set_trace()
        """
        SARSA Update
        Q(s,a) <- Q(s,a) + alpha * (reward + gamma * Q(s',a') - Q(s,a))
        """

别的地方几乎都是一样的，唯一的区别就是learn这个函数，这是唯一的区别。
下面是两个算法的结果：

SARSA

Q-Learning

可以发现一下几点：

1. Q-learning收敛稍稍比SARSA快一点，至少在这个测试环境中
2. 实际上Q-Learning收敛到13步（最优解）走到目标位置，而SARSA收敛到17步
3. Q-Learning比起SARSA在收敛之后，明显更加容易“跌入悬崖”，这个问题我也没有想明白，欢迎讨论，但是可以decay epsilon可以有解决。