Target Networks

What Are Target Networks?

Imagine trying to hit a bullseye on a target that moves every time you throw a dart. No matter how good your aim, the constant movement makes consistent improvement nearly impossible. Now imagine the target stays fixed for 100 throws, then moves to a new position. Suddenly, you can actually practice and improve between resets. Target networks apply this same principle to Q-learning: instead of letting the regression target shift with every gradient update, we freeze a copy of the network and use it to compute stable targets for thousands of steps before refreshing it.

In standard Q-learning with function approximation, the TD target $r + \gamma \max_{a'} Q(s', a'; \mathbf{w})$ depends on the same weights $\mathbf{w}$ being updated. This creates a feedback loop: each weight update changes both the prediction and the target, causing oscillation or divergence. Target networks break this loop by using a separate, periodically updated copy $\mathbf{w}^-$ for the target computation.

How It Works

The Instability Problem

Consider the Q-learning update with function approximation. At each step, we minimize:

$L(\mathbf{w}) = \mathbb{E}\left[\left(y - Q(s, a; \mathbf{w})\right)^2\right]$

where the target is $y = r + \gamma \max_{a'} Q(s', a'; \mathbf{w})$. The gradient is:

$\nabla_\mathbf{w} L = -\mathbb{E}\left[\left(y - Q(s, a; \mathbf{w})\right) \nabla_\mathbf{w} Q(s, a; \mathbf{w})\right]$

The problem: $y$ itself depends on $\mathbf{w}$. When we update $\mathbf{w}$ to make $Q(s, a; \mathbf{w})$ closer to $y$, we simultaneously change $y$. In supervised learning, labels are fixed; in Q-learning with function approximation, the “labels” move with every gradient step. This is analogous to a dog chasing its own tail – the optimization target recedes as the agent approaches it, and generalization in the neural network means that updating one state’s value can unpredictably shift values for other states.

Hard Target Updates

DQN’s original approach (Mnih et al., 2015) uses hard updates: maintain a target network $Q(s, a; \mathbf{w}^-)$ that is an exact copy of the online network, refreshed every $C$ steps:

$\mathbf{w}^- \leftarrow \mathbf{w} \quad \text{every } C \text{ steps}$

Between updates, $\mathbf{w}^-$ is completely frozen. The TD target becomes:

$y = r + \gamma \max_{a'} Q(s', a'; \mathbf{w}^-)$

This target is stationary for $C$ consecutive gradient steps, converting the RL problem into a sequence of supervised regression problems (each with a fixed target for $C$ steps). DQN uses $C = 10{,}000$, meaning the target network is refreshed roughly every 40,000 environment frames.

Soft (Polyak) Updates

An alternative introduced by Lillicrap et al. (2016) in DDPG uses soft updates (also called Polyak averaging):

$\mathbf{w}^- \leftarrow \tau \mathbf{w} + (1 - \tau) \mathbf{w}^-$

This is applied after every gradient step, with $\tau \ll 1$ (typically $\tau = 0.005$ or $\tau = 0.001$). The target network slowly tracks the online network, providing smoother transitions than the abrupt hard update. The target weights become an exponential moving average of online weights.

Recommended visual: Plot comparing the target network weight trajectory under hard updates (staircase function) vs. soft updates (smooth exponential tracking) over training steps.

Why This Works: Stabilizing the Optimization Landscape

Target networks convert a non-stationary optimization problem into a series of approximately stationary ones. Between target updates, the loss surface $L(\mathbf{w})$ has a fixed shape because the targets $y$ are constant. The optimizer can make genuine progress toward the current target before the landscape shifts.

Formally, the Bellman operator $\mathcal{T}$ is a $\gamma$-contraction in the tabular case:

$\|\mathcal{T}Q_1 - \mathcal{T}Q_2\|_\infty \leq \gamma \|Q_1 - Q_2\|_\infty$

Target networks approximate applying $\mathcal{T}$ to a fixed $Q^-$ and then fitting $Q$ to the result, which more closely mirrors the provably convergent iteration $Q_{k+1} = \mathcal{T}Q_k$ than the fully online alternative where $Q$ and the target change simultaneously.

Why It Matters

Target networks are one of the two innovations (alongside experience replay) that made deep Q-learning practical. Without them, the Q-network’s loss landscape shifts erratically with each weight update, causing training to oscillate or diverge. The technique is now ubiquitous: virtually every value-based deep RL algorithm uses target networks, and the concept has been extended to actor-critic methods (DDPG, TD3, SAC) where both the critic and sometimes the actor have target copies.

Key Technical Details

• Hard update frequency: DQN uses $C = 10{,}000$ gradient steps. Values too small (e.g., $C = 100$) provide insufficient stability; values too large (e.g., $C = 100{,}000$) cause the target to lag excessively behind the learned value.
• Soft update rate: $\tau = 0.005$ is standard for continuous-control algorithms (DDPG, TD3, SAC). Larger $\tau$ values track faster but provide less stability.
• Hard vs. soft: Hard updates are simpler but create discontinuities when the target network suddenly jumps. Soft updates are smoother but add a hyperparameter. In practice, soft updates are preferred for continuous-action domains; hard updates remain common for discrete-action domains.
• Memory cost: Target networks double the model’s parameter memory. For DQN’s ~1.7M parameters this is negligible; for large models it can matter.
• No gradient through targets: The target $y = r + \gamma \max_{a'} Q(s', a'; \mathbf{w}^-)$ is treated as a constant during backpropagation. Gradients flow through $Q(s, a; \mathbf{w})$ only.
• Initialization: The target network is initialized as an exact copy of the online network at the start of training.

Common Misconceptions

• “Target networks eliminate instability entirely.” They reduce it substantially but do not guarantee convergence. The combination of function approximation, bootstrapping, and off-policy data (the deadly triad from Function Approximation) remains a fundamental challenge. Target networks make training practical, not provably stable.

• “Soft updates are always better than hard updates.” This depends on the domain. Hard updates with appropriate $C$ work well for discrete-action problems (Atari). Soft updates are preferred in continuous control. There is no universal winner.

• “The target network learns slower on purpose.” The target network does not learn at all – it is passively updated (copied or blended) from the online network. Only the online network receives gradient updates. The target network is a reference point, not a learner.

Connections to Other Concepts

Further Reading