What is the Difference Between Q Learning and Sarsa?

What is the Difference Between Q Learning and Sarsa?

Artificial Intelligence & Machine Learning

Q-Learning vs SARSA

In the realm of reinforcement learning, Q-Learning and State-Action-Reward-State-Action (SARSA) are two pivotal algorithms that have shaped our understanding and application of learning in environments where an agent must make decisions sequentially. This essay aims to provide a detailed comparison of Q-Learning and SARSA, including their theoretical underpinnings, calculation methods, and real-world applications.

Background and Calculation of Q-Learning

Introduction to Q-Learning: Q-Learning is a model-free reinforcement learning algorithm that aims to learn a policy, dictating the best action to take in a given state to maximize cumulative reward. It was introduced by Christopher Watkins in 1989 and has been a foundational method in various fields requiring decision-making strategies.

Calculation in Q-Learning: The Q-value function in Q-Learning, Q(s, a), estimates the value of taking action ‘a’ in state ‘s’. The algorithm updates the Q-values using the Bellman equation:

Diagram showing the components in a typical Reinforcement Learning (RL) system. An agent takes actions in an environment which is interpreted into a reward and a representation of the state which is fed back into the agent. Incorporates other CC0 work: https://openclipart.org/detail/202735/eye-side-view https://openclipart.org/detail/191072/blue-robot and https://openclipart.org/detail/246662/simple-maze-puzzle

Real-World Application of Q-Learning: Q-Learning is widely used in automated control systems, robotics, and gaming, where the environment can be modeled as a Markov Decision Process (MDP).

Background and Calculation of SARSA

Introduction to SARSA: SARSA, introduced by Rummery and Niranjan in 1994, is another model-free reinforcement learning algorithm. Unlike Q-Learning, SARSA is an on-policy method, meaning it learns the value of the policy being followed, including the exploration steps.

Calculation in SARSA: In SARSA, the update rule is slightly different from Q-Learning, incorporating the next action:



  • �′a′ is the action taken in the new state �′s′.
  • Other terms are as defined in the Q-Learning section.

Real-World Application of SARSA: SARSA is useful in environments where the learning policy itself affects the outcome, such as in certain control systems or scenarios where safety during learning is a concern.

Comparison: Q-Learning vs SARSA

  1. Exploration Strategies:
    • Q-Learning is an off-policy learner, meaning it learns the value of the optimal policy independently of the agent’s actions.
    • SARSA is an on-policy learner, learning the value of the policy being followed, including the exploratory moves.
  2. Convergence Behavior:
    • Q-Learning tends to be more aggressive in its learning approach, directly approximating the optimal policy.
    • SARSA takes a more conservative approach, considering the consequences of exploratory actions, which can lead to safer policy learning.
  3. Stability and Performance:
    • Q-Learning can sometimes diverge when function approximation is involved.
    • SARSA is generally more stable but might converge to a suboptimal policy due to the inclusion of exploratory actions in its updates.
  4. Real-world Implications:
    • Q-Learning is better suited for scenarios where the optimal policy needs to be learned without much concern for the risks involved in exploration.
    • SARSA is more appropriate for tasks where the safety and the costs of exploration are significant, such as autonomous driving or robotic movement in sensitive environments.


Both Q-Learning and SARSA have their unique strengths and are better suited for different types of problems. Q-Learning’s ability to learn the optimal policy aggressively makes it ideal for scenarios where exploration risks are low. In contrast, SARSA’s conservative approach is beneficial in environments where the cost of exploration is high, and safety is a priority. Understanding the specific requirements and constraints of the problem at hand is crucial in choosing between these two powerful reinforcement learning algorithms.



State–action–reward–state–action – Wikipedia

What is the Difference Between Q Learning and Sarsa? What is the Difference Between Q Learning and Sarsa?