Explainable Reinforcement Learning for Adaptive Traffic Signals

Explainable Reinforcement Learning for Adaptive Traffic Signal Control, an arXiv paper by Dickens Kwesiga, Nishu Choudhary, Angshuman Guin and Michael Hunter, was submitted on 4 Jul 2026 as arXiv:2607.03703 (cs.AI). The paper presents a novel, entity-centric reinforcement learning framework for traffic signal control that emphasizes explicit interpretability and operational safety.

What did the authors build?

The paper introduces an entity-centric RL architecture that disaggregates intersection observations into high-dimensional lane entities and phase temporal configurations, producing a real time affinity matrix that quantifies the influence of signal phases on approach volumes and queues. The design combines this representation with a deterministic action-masking interface integrated directly into the Proximal Policy Optimization pipeline to block invalid phase transitions and enforce signal timing and safety constraints.

The authors evaluated the system in a microscopic simulation environment and report that it outperforms state-of-the-art baselines in delay minimization. The submission is available on arXiv with DOI https://doi.org/10.48550/arXiv.2607.03703.

How does the explainability work?

Explainability is provided by a dual-stage attention network featuring sequential multi-head cross-attention and self-attention blocks that dynamically extract relational dependencies and inter-lane conflicts. That attention stack produces an affinity matrix which the authors describe as providing "full visual and analytical interpretability." The network thus links signal phases to specific approach volumes and queues, so attention weights can be inspected and audited.

Structurally, the paper departs from flat state vectors: it preserves the intersection topology by treating lanes and phase timings as distinct entities. The cross-attention captures interactions between lanes and phases; self-attention models intra-entity relations. The resulting attention weights, the authors report, align with established traffic engineering principles, offering what they call an auditable and trust-enabling architecture.

How does it ensure safety and operational reliability?

Safety is enforced by a deterministic action-masking interface that is integrated into the Proximal Policy Optimization pipeline to explicitly block invalid phase transitions. The mask guarantees compliance with established signal timing and safety constraints by preventing the agent from selecting prohibited actions at runtime. This deterministic masking is paired with the learned policy so that the RL agent cannot execute unsafe or illegal phase changes.

In addition to the mask, the entity-centric input preserves geometric and topological structure so that relational reasoning about conflicts and priorities remains visible through attention weights rather than buried in opaque latent vectors.

Why it matters

Bringing explicit interpretability into adaptive traffic control addresses barriers that often prevent transportation agencies from adopting deep RL: regulatory compliance, operational trust, and the need for auditable behavior. The paper couples interpretability with an explicit safety mechanism, the deterministic action mask, which directly tackles the operational reliability requirement of signal control systems. If attention-based explanations truly align with traffic engineering principles and the controller reduces delay versus baselines, that combination could make trial deployments easier to justify.

What to watch

Look for follow-up work or code and simulation details the authors may release: the claim that the approach "outperforms state-of-the-art baselines in delay minimization" rests on microscopic simulation results stated in the submission. Practical next steps that will matter are open-source evaluation artifacts, comparisons on standardized traffic benchmarks, and any real-world pilot where the deterministic mask and attention visualizations are used by traffic engineers.

References and specifics: the paper was submitted 4 Jul 2026, is listed as arXiv:2607.03703 [cs.AI], and links to the DOI https://doi.org/10.48550/arXiv.2607.03703. The authors name their core components as an entity-centric representation, a dual-stage attention network with sequential multi-head cross-attention and self-attention blocks, a real time affinity matrix for interpretability, and a deterministic action-masking interface inside Proximal Policy Optimization. The evaluation is reported in a microscopic simulation environment and the authors state improved delay minimization compared with state-of-the-art baselines.