SARSA Reinforcement Learning

Prerequisites: Q-Learning technique

SARSA algorithm is a slight variation of the popular Q-Learning algorithm. For a learning agent in any Reinforcement Learning algorithm it’s policy can be of two types:-

On Policy: In this, the learning agent learns the value function according to the current action derived from the policy currently being used.
Off Policy: In this, the learning agent learns the value function according to the action derived from another policy.

Q-Learning technique is an Off Policy technique and uses the greedy approach to learn the Q-value. SARSA technique, on the other hand, is an On Policy and uses the action performed by the current policy to learn the Q-value.

This difference is visible in the difference of the update statements for each technique:-

Q-Learning: $Q(s_{t},a_{t}) = Q(s_{t},a_{t}) + \alpha (r_{t+1}+\gamma max_{a}Q(s_{t+1},a)-Q(s_{t},a_{t}))$
SARSA: $Q(s_{t},a_{t}) = Q(s_{t},a_{t}) + \alpha (r_{t+1}+\gamma Q(s_{t+1},a_{t+1})-Q(s_{t},a_{t}))$

Here, the update equation for SARSA depends on the current state, current action, reward obtained, next state and next action. This observation lead to the naming of the learning technique as SARSA stands for State Action Reward State Action which symbolizes the tuple (s, a, r, s’, a’).

The following Python code demonstrates how to implement the SARSA algorithm using the OpenAI’s gym module to load the environment.

Step 1: Importing the required libraries

import numpy as np 
import gym 

Step 2: Building the environment

Here, we will be using the ‘FrozenLake-v0’ environment which is preloaded into gym. You can read about the environment description here.

#Building the environment 
env = gym.make('FrozenLake-v0') 

Step 3: Initializing different parameters

#Defining the different parameters 
epsilon = 0.9
total_episodes = 10000
max_steps = 100
alpha = 0.85
gamma = 0.95
  
#Initializing the Q-matrix 
Q = np.zeros((env.observation_space.n, env.action_space.n)) 

Step 4: Defining utility functions to be used in the learning process

#Function to choose the next action 
def choose_action(state): 
    action=0
    if np.random.uniform(0, 1) < epsilon: 
        action = env.action_space.sample() 
    else: 
        action = np.argmax(Q[state, :]) 
    return action 
  
#Function to learn the Q-value 
def update(state, state2, reward, action, action2): 
    predict = Q[state, action] 
    target = reward + gamma * Q[state2, action2] 
    Q[state, action] = Q[state, action] + alpha * (target - predict) 

Step 5: Training the learning agent

#Initializing the reward 
reward=0
  
# Starting the SARSA learning 
for episode in range(total_episodes): 
    t = 0
    state1 = env.reset() 
    action1 = choose_action(state1) 
  
    while t < max_steps: 
        #Visualizing the training 
        env.render() 
          
        #Getting the next state 
        state2, reward, done, info = env.step(action1) 
  
        #Choosing the next action 
        action2 = choose_action(state2) 
          
        #Learning the Q-value 
        update(state1, state2, reward, action1, action2) 
  
        state1 = state2 
        action1 = action2 
          
        #Updating the respective vaLues 
        t += 1
        reward += 1
          
        #If at the end of learning process 
        if done: 
            break

In the above output, the red mark determines the current position of the agent in the environment while the direction given in brackets gives the direction of movement that the agent will make next. Note that the agent stays at it’s position if goes out of bounds.

Step 6: Evaluating the performance

#Evaluating the performance 
print ("Performace : ", reward/total_episodes) 
  
#Visualizing the Q-matrix 
print(Q) 

My Personal Notes

Recommended Posts: