A new add-on layer for deep reinforcement learning networks makes Q-learning converge faster — without touching the underlying algorithm.
The paper introduces a "mean-expansion layer" that changes how action-values are learned inside Q-networks. Instead of updating each state-action pair in isolation, the layer shares value information across all actions available in a given state. It also reframes the learning target: rather than estimating potentially large raw action-values directly, the network learns a lower-norm representation of them, which is easier to move from a near-zero initialization toward the true value. The approach is parameter-free, meaning it slots into existing architectures as a structural change, not extra weights to train.
The results land on a well-worn benchmark: 57 Atari games, the industry's standard stress test for RL agents. Applied to both deep Q-networks and implicit quantile networks, the layer improved aggregate performance across that suite while widening action gaps — the margin by which an agent prefers its best move — and sharply reducing value overestimation, a known failure mode where Q-networks inflate their confidence.
Value overestimation has dogged Q-learning since DeepMind first paired it with deep networks in 2013; the standard fix, Double DQN, addressed it with a second network rather than a structural layer. A parameter-free patch that reduces overestimation at the architecture level is a different angle — though whether it holds outside Atari grids, in continuous-action or real-world environments, is the question the paper leaves open.