While using deep neural networks to approximate the action-value function, $Q(s, a; \theta)$ , seems like a straightforward extension of the function approximation techniques discussed previously, applying them directly within the Q-learning framework introduces significant hurdles. Standard supervised learning methods, where neural networks excel, often rely on assumptions that don't hold true in the reinforcement learning setting. Training stability becomes a major concern due to two primary factors: the sequential nature of the data and the constantly changing target values.

The Problem of Correlated Data

In typical supervised learning scenarios, training data consists of samples assumed to be independent and identically distributed (i.i.d.). This independence allows optimization algorithms like stochastic gradient descent (SGD) to make steady progress by sampling mini-batches that provide relatively unbiased estimates of the true gradient.

Reinforcement learning, however, generates data sequentially. An agent interacts with the environment over time, producing a sequence of experiences: $(s_t, a_t, r_{t+1}, s_{t+1}), (s_{t+1}, a_{t+1}, r_{t+2}, s_{t+2}), \dots$ . Consecutive samples in this sequence are often highly correlated. The state $s_{t+1}$ strongly depends on $s_t$ and $a_t$ . If we train the neural network on these experiences as they arrive, we violate the i.i.d. assumption.

Why is this correlation detrimental?

Inefficient Learning: Training on highly correlated samples can lead to inefficient learning. The network might overfit to recent experiences, forgetting what it learned from past, potentially different, situations. Imagine learning to play a game by only practicing the first level repeatedly; you might become very good at that level but fail completely on level two. Correlated updates can push the network parameters in a direction specific to a small, temporally-linked region of the state-action space, potentially undoing previous learning.
Gradient Variance: While correlated samples don't necessarily bias the gradient estimate in the long run, they can significantly increase its variance within mini-batches. This makes the training process noisier and potentially slower to converge.

The Instability of a Moving Target

The second major challenge arises from the way Q-learning updates its estimates. Recall the target value used in the Q-learning update for a transition $(s, a, r, s')$ :

\text{Target} = r + \gamma \max_{a'} Q(s', a'; \theta)

Here, $\theta$ represents the parameters of our neural network approximating the Q-function. We want to adjust the network's prediction $Q(s, a; \theta)$ to be closer to this target value. The loss function, often the Mean Squared Error (MSE), would look something like:

L(\theta) = \mathbb{E} \left[ \left( \underbrace{r + \gamma \max_{a'} Q(s', a'; \theta)}_{\text{Target}} - \underbrace{Q(s, a; \theta)}_{\text{Prediction}} \right)^2 \right]

Notice that the parameters $\theta$ appear in both the prediction $Q(s, a; \theta)$ and the target $r + \gamma \max_{a'} Q(s', a'; \theta)$ . When we perform a gradient descent step to minimize this loss, we are adjusting $\theta$ . But as $\theta$ changes, the target value itself shifts because it also depends on $\theta$ .

This creates a "moving target" problem. We are trying to make our network's predictions match a target that is also changing with every update step. This coupling can lead to feedback loops and instability: an update aiming to reduce the error might inadvertently shift the target in a way that increases the error in the next step, potentially causing oscillations or even divergence of the network parameters. It's like trying to aim at a target that jerks away every time you adjust your aim based on your last shot.

These two issues, correlated data and the non-stationary target, mean that naively combining standard Q-learning with deep neural networks often fails to converge or produces unstable results. Addressing these challenges is fundamental to making deep reinforcement learning practical, leading to the development of techniques like Experience Replay and Fixed Q-Targets, which we will explore next.