为什么强大的AI仍难以接管城市交通？IJCAI 2026给出三点答案 Why Powerful AI Still Cannot Run City Traffic: IJCAI 2026's Answer
交通AI的模型和分数越来越强，城市却没有因此大规模智能化。IJCAI 2026相关分享指出，真正卡住落地的不是单一算法，而是缺失数据、跨城迁移、现实约束和可解释性同时过关的问题。 Traffic AI keeps improving in model size and benchmark scores, yet real cities still have not scaled it. An IJCAI 2026 talk argues the bottleneck is not one algorithm, but the combined challenge of missing data, cross-city transfer, real-world constraints, and explainability.
背景
过去几年，交通AI研究不断追求更大的模型、更复杂的结构和更高的测试集分数，但真实城市并没有因此大规模实现智能化。2023年麦肯锡针对工业和交通领域的调研显示，受访者普遍认为距离真正智能化仍有10到20年。
为什么学术模型难落地
- 学术研究常用的是整理得很干净的数据，公共部门面对的却是缺失、混乱、实时变化的数据。
- 有的超大城市只有少数路口安装传感器，很多区域没有直接观测。
- 研究中会通过遮掉部分节点、加入虚拟节点等方法，推断缺失位置的交通状态。
- 信号灯控制还要面对跨城迁移、多个路口协同，以及修路、封桥、马拉松等新情形的快速适应。
真正的门槛
在强化学习信号灯控制上，团队还尝试加入更符合交通工程习惯的约束，让模型少做一些工程师难以理解的动作。但全球目前仍没有城市长期部署基于强化学习的信号灯系统，原因包括解释不清、仿真环境与现实设备不匹配，以及未来互联车辆普及后是否还需要集中控制器等问题。
PDE三角必须同时成立：感知、决策和解释性都要过关。
四步框架
- 先说明发生了什么
- 再找出为什么发生
- 接着预测将要发生什么
- 最后给出应该怎么做
Background
Over the past few years, traffic AI research has kept pushing for larger models, more complex architectures, and higher benchmark scores, yet real cities have not scaled intelligence in the same way. A 2023 McKinsey survey in industrial and transport domains suggested respondents still see 10 to 20 years before true intelligence becomes reality.
Why the models do not land in practice
- Academic work often uses clean, curated data, while public agencies deal with missing, messy, and constantly changing data.
- In some very large cities, only a small number of intersections have sensors, leaving much of the network without direct observation.
- The research described here uses techniques such as masking nodes during training and adding virtual nodes to infer traffic states where data is absent.
- Signal control also has to handle cross-city transfer, coordination across intersections, and fast adaptation to roadworks, bridge closures, marathons, and other novel events.
The real bottleneck
For reinforcement-learning-based signal control, the team also added constraints that better match traffic engineering practice, so the model makes fewer moves that engineers cannot easily understand. Still, no city has long-term deployed a reinforcement-learning traffic-light system at scale. The cited reasons include weak explainability, a gap between simulation and real hardware, and open questions about whether centralized control will still be needed once connected vehicles become widespread.
The PDE triangle has to hold at the same time: perception, decision-making, and explainability.
The four-step frame
- Describe what happened
- Find out why it happened
- Predict what will happen next
- Then decide what should be done
来源
- 雷锋网 · 09-01 10:54