Question $4$

Reinforcement Learning

$[8$ Marks$]$

Explain how Q$-$learning overcomes the challenge of having to act greedily with

respect to a value function.

Describe what is meant by the exploration$-$exploitation dilemma.

Write down the SARSA update rule. How does this differ from the Q$-$learning

update rule?

What is the main difference between early $($pre $2000)$ attempts at function approx$-$

imation, and function approximation using deep learning $($with neural networks?$)$

$[1]$

Describe how the DQN algorithm overcomes the problem training using data that

is highly correlated.