PIXELBANKv8.2.1
Menu
Back to RL Study Plan
Week 5

Chapter 5: Q-Learning & SARSA

Model-free control with action values: SARSA as on-policy TD control, Q-learning as its off-policy counterpart, what the on-policy/off-policy distinction actually costs, and the maximization-bias fixes of Expected SARSA and Double Q-learning.

Chapter Overview

Chapter 4 built the machinery for learning values from experience. This chapter turns it into control. The move is small — learn Q(s,a)Q(s,a) instead of V(s)V(s) so that improvement needs no model — but it produces the two algorithms that defined reinforcement learning for two decades.

SARSA and Q-learning differ by a single term in the update, and that one difference is the on-policy/off-policy distinction. SARSA learns the value of the policy it is actually following, including its exploratory mistakes. Q-learning learns the value of the greedy policy regardless of what it is doing. On the cliff-walking problem they produce visibly different behaviour: one takes the safe route, the other the optimal-but-risky one. Understanding why is understanding the distinction that governs everything from replay buffers to importance sampling.

Chapter Roadmap

Click any topic to jump in

1
Action-Value Methods

Learning Q rather than V makes greedy improvement possible without any model of the environment.

Two ways to form the bootstrap target

Use the action you will actually take, or the best action available — one choice, two algorithms.

2
SARSA

On-policy TD control: bootstrap from the action actually taken next, so exploration is priced into the value.

3
Q-Learning

Off-policy TD control: bootstrap from the best next action, learning the greedy policy while behaving otherwise.

That choice has a name and consequences
4
On vs Off-Policy

Whether the policy being evaluated is the one generating the data — the distinction that decides if replay is legal.

Both targets can be improved
5
Expected SARSA & Double Q

Reduce update variance by averaging over actions, and remove maximization bias with two decoupled estimators.

Sign up to unlock this chapter

This chapter is part of PixelBank Premium. Create a free account, then upgrade to read the full lesson — concepts, walkthroughs, and exercises.