Communication-Aware Adaptive Decentralized Traffic Signal Control via Broadcast gate

Abstract

GNN-based multi-agent reinforcement learning (MARL) for adaptive traffic signal control typically assumes unrestricted neighbor communication at every decision step, with the cost of that messaging rarely measured. We propose a learned sparse communication gating framework in which each intersection decides, via a differentiable Gumbel-Sigmoid gate trained jointly with the Q-network under a sparsity penalty $\alpha$, whether to broadcast its local state. For fair comparison we define an effective neighbor communication ratio $\rho_{\text {eff }}$ and pin saturated-communication baselines (CoLight, MAPPO, Alwayson) at $\rho_{\mathrm{eff}}=1$. On a $10 \times 10$ SUMO grid, two findings emerge. Under rich pressure observations, the gate at $\alpha=0.05$ attains the lowest mean total waiting time among tested methods (13,886 s over five seeds) with paired bootstrap 95% CIs of the per-seed difference excluding zero against Always-on, CoLight-feat, and CoLight-attn on the original three-seed subset, while broadcasting at $\rho_{\text {eff }} \approx 5 \times 10^{-4}$; the same near-zero regime is robust across $\alpha \in[0.05,0.30]$. Under partial scalar observations with a tight broadcast budget, the same gate collapses to a similar $\rho_{\text {eff }}$ for every $\alpha$ but underperforms the saturated baselines-an honest negative finding that motivates observation-quality-conditioned gating rather than $\alpha$ tuning. A learning-time trajectory shows $\rho_{\text {eff }}$ decaying from the Gumbel prior of $\approx 0.5$ in lockstep with waiting time, ruling out a measurement artifact. Together, the two regimes argue that communication necessity in traffic MARL is not an architectural constant but a function of observation-reward redundancy.