Planning Around Shock Births: Event-Split Search for Moving-Bottleneck Traffic Control

A single automated car driving slowly in heavy traffic acts as a moving bottleneck. Cars pile up behind it and spread out ahead of it. That gives traffic engineers a lever: by choosing the car's speed, one can deliberately reshape congestion waves and dissolve stop-and-go traffic. In the mathematical model, the pile-up behind the car is a "non-classical shock," a sudden drop in density that ordinary traffic physics forbids. It appears only when the car's capacity limit is binding, and it vanishes when traffic thins out. Planning the car's speed over tens of kilometres and tens of minutes is hard for a specific reason. The cost of a speed plan changes smoothly while the set of binding capacity limits stays fixed, and it jumps or kinks whenever that set changes. Flat optimizers, which treat the whole horizon as one problem, stall on exactly these kinks.

We propose a divide-and-conquer planner that cuts the horizon at the moments when the car's capacity constraint switches on or off, which are the moments when a non-classical shock is born or dies. Generic recursive planners split at temporal midpoints. Our planner does not; it searches over the order of these switching events and solves only smooth subproblems between them. The proposal has a second half. These switching times are the quantity most corrupted by numerical dissipation, the artificial smoothing that coarse simulators add. If the simulator that scores a plan smears the density behind the car, it moves the switch times, and the planner optimizes against a fictional landscape. We will measure that bias directly. We will also show that the planner must be paired with a scheme that keeps the jump sharp and conserves the number of vehicles.

Core Claim

Long-horizon Lagrangian control means steering bulk traffic by controlling one vehicle inside it. We claim this problem is a hybrid problem with a small discrete skeleton: the sequence of constraint activations and deactivations, and which upstream waves the vehicle engages in which order. Between consecutive switches the dynamics are smooth. Making those switches the split points of a recursive AND/OR search turns one non-convex long-horizon optimization into a tree search over a short discrete schedule, plus local gradient solves on smooth segments. Finite wave speed is what makes this tractable. A subgoal does not need to specify the whole density field, only the vehicle's state and the few wave fronts inside its domain of influence.

We make a second, falsifiable claim. Plans scored with a dissipative evaluator, such as first-order Godunov on a grid coarse enough to be affordable inside a search tree, will systematically overestimate the benefit of control. Part of the "dissolved" congestion was dissolved by the numerics. That optimism should shrink under grid refinement and disappear with a conservative reconstruction scheme.

Research Program

The setting is the coupled model of Modeling of moving bottlenecks in traffic flow: a PDE-ODE model with moving density constraints, whose solutions are shown to exist in Scalar conservation laws with moving constraints arising in traffic flow modeling: an existence result. Traffic density $\rho(x,t)$ obeys a scalar conservation law. The bottleneck position $y(t)$ obeys an ODE. The flux relative to the bottleneck is capped, $f(\rho) - \dot y \rho \le F_{\max}(\dot y)$. We write the control objective as a sum over segments delimited by switching times $\tau_0 < \tau_1 < \dots < \tau_K$, where the cap changes between binding and slack:

$$J(v) = \sum_{k=1}^{K} J_k\big(v|_{[\tau_{k-1},\tau_k]};\ \sigma_{k-1}\big),$$

Here $v$ is the vehicle's speed command and $\sigma_{k-1}$ is the local wave configuration at $\tau_{k-1}$: the vehicle's position plus the positions and states of the shocks inside its forward domain of influence. Each $J_k$ is smooth in $v$ when the active set is fixed. All the non-smoothness of $J$ sits in the choice of $K$ and the event ordering.

The planner reuses the AND/OR structure of Divide-and-Conquer Monte Carlo Tree Search For Goal-Directed Planning. That paper proposes midpoint subgoals with a learned prior, scores sub-tasks with a learned reachability value, and trains both from hindsight midpoints of successful trajectories. We keep the tree and change three things. First, candidate split points are predicted switching events rather than temporal midpoints. They are proposed by computing when the upstream shock that the vehicle is approaching reaches it, a calculation that follows from the jump (Rankine–Hugoniot) speeds of the shocks. Second, a subgoal $\sigma$ is the low-dimensional wave configuration rather than a full state. This addresses the continuous-branching weakness that the DC-MCTS authors acknowledge. Third, the reachability value of a sub-task is checked against the cone of possible wave propagation before any learned estimate is consulted. A subgoal that would require a wave to outrun its maximum characteristic speed is pruned. This targets the failure of physics-agnostic subgoals noted in the research file. Inside each leaf segment, a short-horizon gradient MPC of the kind used in Lagrangian Control of Conservation Laws and Mixed Models optimizes speed with the active set held fixed.

Leaves are evaluated with the conservative fixed-grid scheme of A conservative scheme for non-classical solutions to a strongly coupled PDE-ODE problem. That scheme reconstructs the sub-cell jump at the bottleneck so the cap is enforced exactly and vehicle count is conserved. For multiple interacting vehicles we use the constrained-Riemann-solver scheme of Interacting moving bottlenecks in traffic flow. That paper reports small spurious oscillations precisely when bottlenecks switch between active and inactive. Because our planner splits exactly there, those oscillations are a direct threat. We will quantify how much they perturb predicted switch times before trusting multi-vehicle results.

We deliberately do not use a learned flux as the evaluator. Supervised and Unsupervised Neural Network Solver for First Order Hyperbolic Nonlinear PDEs learns a conservative finite-volume flux that matches discontinuous Galerkin accuracy on 2,000 Riemann problems. Data-driven optimized high-order WENO schemes with low-dissipation and low-dispersion learns corrections that reduce WENO's low-pass behaviour. Both were trained only on classical, entropy-satisfying shocks. The non-classical jump is exactly the kind of discontinuity those schemes are built to remove. A learned evaluator that respects the moving cap is a possible follow-on; it is not a premise of this work.

Baselines

The naive reference is no control. The automated vehicle drives at its desired free-flow speed and slows only when the model forces it to. The same initial congestion profiles, solver and road length are used. Every controlled method is reported as a relative improvement over this run.

The strongest existing method is flat receding-horizon MPC on the moving-bottleneck model, as in Lagrangian Control of Conservation Laws and Mixed Models. On a simulated 50 km three-lane highway it reports approximately 2.05% lower total fuel consumption and 5.8% lower average travel time. We will reimplement it and first reproduce numbers in that range on a matched scenario. Only then will we run it on the same leaf evaluator, the same fuel model and the same wall-clock planning budget per control step as our planner, with its horizon and step size tuned by grid search under that budget. Beating it means lower fuel and travel time than flat MPC at equal compute on horizons of 50 km and longer, with the margin growing as the horizon grows.

The second competitor is multiple shooting with uniform splits. It cuts the horizon into the same number of segments our planner uses, at evenly spaced times, and optimizes them jointly with continuity constraints. It serves as the ablation for our key ingredient: it has decomposition but no event alignment. Generic DC-MCTS, with learned midpoint prior and value and trained with hindsight relabelling on the same simulator rollouts, is the ablation that isolates the physics-shaped subgoal space and the cone pruning.

A third ablation isolates the evaluator. It runs our planner unchanged but scores leaves with coarse first-order Godunov at the grid resolution that fits the same budget. All final plans from all methods are re-scored with a fine-grid conservative reference solve, so no method is graded by its own simulator.

Why not just…?

Why not just run flat MPC with a finer grid and more compute? This is the real threat, and we treat it as one. Bayen and Delle Monache's MPC already works at 50 km. If warm-started flat MPC matches our planner at equal compute as the horizon is stretched, the decomposition is unnecessary and we will say so. The horizon-scaling experiment below is designed to settle this. Our prediction rests on the cost landscape being piecewise smooth with kinks at switches. We will check that prediction by plotting the cost along random directions in speed space and counting kinks against detected switch events.

Why not just use generic recursive planning such as DC-MCTS with temporal midpoints? The DC-MCTS paper itself identifies inaccurate reachability values and large branching over continuous subgoals as its weak points. A full density profile is a very high-dimensional subgoal. The generic-DC-MCTS baseline measures how often its subgoals are infeasible under the reference solver and how much compute it spends on them.

Why not just train a flat PPO policy in a microscopic simulator, as in the mixed-autonomy work? Traffic Smoothing Controllers for Autonomous Vehicles Using Deep Reinforcement Learning and Real-World Trajectory Data reports over 15% fuel savings at 4% penetration using car-following simulation calibrated on I-24 data. That result is strong, but it answers a different question: local smoothing by many vehicles. Benchmarks for reinforcement learning in mixed-autonomy traffic reports flat PPO and TRPO being sample-inefficient and finding reward-gaming policies on bottleneck networks. We will not claim superiority across simulators. We will include a PPO policy trained directly on our macroscopic simulator, under the same interaction budget, as a learned-flat reference.

Why not just use a standard shock-capturing scheme inside the search? Standard high-order schemes enforce classical entropy conditions and smooth away the non-classical drop. Coarse first-order schemes converge slowly on these problems. The Chalons et al. conservative scheme itself converges at rates between $O(\Delta x^{0.8})$ and $O(\Delta x^{1.5})$, so cheap grids are far from exact. The evaluator ablation measures how much this matters for decisions, not only for solution error.

Proposed Experiments

We first validate the evaluator on the Cases 0–IV Riemann suite from A conservative scheme for non-classical solutions to a strongly coupled PDE-ODE problem. We check exact vehicle conservation and the reported convergence range. We then add a new measurement, the error in predicted constraint-switch times as a function of grid size, for both the conservative scheme and Godunov.

Control experiments use a periodic ring-road version of the LWR model, with congestion waves seeded to mimic the 22-vehicle setting of the Arizona experiment in Lagrangian Control of Conservation Laws and Mixed Models. They also use highway stretches of 10, 25, 50 and 100 km with one controlled vehicle. Multi-vehicle runs on the same stretches use two to four vehicles, including same-lane queuing as in Interacting moving bottlenecks in traffic flow. Initial density profiles are drawn from synthetic stop-and-go patterns; where possible they are also shaped from macroscopic aggregates of the I-24 trajectories. Because no macroscopic validation numbers appear in the file for I-24 (no reported result found on Lacuna for a macroscopic moving-bottleneck model calibrated to I-24), we treat that as initialization realism, not validation. Metrics are fuel under the Bayen–Delle Monache fuel model, average travel time, planning wall-clock time, and the fraction of proposed subgoals that the reference solver deems infeasible.

Three checks decide the proposal. The first is horizon scaling. At fixed per-step compute, we expect the gap between the event-split planner and flat MPC to widen from 10 km to 100 km. If it does not, the core claim fails. The second is event alignment. Our planner must beat uniform-split multiple shooting with the same number of segments; if it does not, the gain comes from decomposition generally rather than from splitting at shock births and deaths. The third is dissipation bias. For every planner scored with coarse Godunov, we report the predicted improvement minus the improvement under the fine-grid conservative reference. We expect this gap to be positive, to shrink as the grid is refined, and to be near zero for the conservative evaluator. A positive, refinement-dependent gap would show that part of the control benefit reported by dissipative simulators is a numerical artefact. That result would hold whether or not the planner itself wins.