SignalClaw: LLM-Guided Evolutionary Synthesis of Interpretable Traffic Signal Control Skills
Summary: arXiv:2604.05535v1 Announce Type: new
Abstract: Traffic signal control (TSC) requires strategies that are both effective and interpretable for deployment, yet reinforcement learning produces opaque neural policies while program synthesis depends on restrictive domain-specific languages. We present SIGNALCLAW, a framework that uses large language models (LLMs) as evolutionary skill generators to synthesize and refine interpretable control skills for adaptive TSC. Each skill includes rationale, selection guidance, and executable code, making policies human-inspectable and self-documenting.
Overview of SignalClaw
SignalClaw is an innovative framework designed to enhance traffic signal control systems by integrating advanced artificial intelligence techniques. By harnessing the power of large language models, SignalClaw creates interpretable traffic signal control skills that can adapt to varying traffic conditions.
Key Features
- Skill Generation: SignalClaw employs LLMs to generate and refine control skills, ensuring that each skill is understandable and can be easily adjusted by traffic engineers.
- Evolutionary Feedback: The framework translates simulation metrics such as queue percentiles and delay trends into natural language feedback, which guides the evolution of traffic control skills.
- Event-Driven Compositional Evolution: An event detector identifies various traffic scenarios, including emergencies and congestion, allowing for quick adaptation of traffic signals.
- Independent Skill Evolution: Each skill can evolve independently, and a priority chain enables the composition of skills without the need for retraining.
Performance Evaluation
SignalClaw was evaluated in both routine and event-injected scenarios using the Simulation of Urban MObility (SUMO) tool. The results indicated a significant improvement in traffic management:
- In routine scenarios, SignalClaw achieved average delays ranging from 7.8 to 9.2 seconds, performing within 3 to 10 percent of the best existing methods while maintaining low variance across random seeds.
- During event scenarios, SignalClaw demonstrated the lowest emergency delay, achieving results between 11.2 to 18.5 seconds, compared to 42.3 to 72.3 seconds for the MaxPressure method and 78.5 to 95.3 seconds for DQN.
- For transit delays, the framework recorded delays of 9.8 to 11.5 seconds, significantly outperforming MaxPressure, which recorded delays of 38.7 to 45.2 seconds.
Conclusion
SignalClaw represents a significant advancement in traffic signal control technology, combining the interpretability of traditional methods with the adaptive capabilities of modern AI. The ability to evolve skills from simple linear rules to complex conditional strategies ensures that traffic engineers can effectively manage urban traffic systems. This innovation not only improves traffic flow but also enhances overall safety by prioritizing emergency vehicles and responding to real-time traffic incidents.