Mixed Traffic Control Network

Updated 14 December 2025

Mixed Traffic Control Networks are urban/suburban systems where human-driven and autonomous vehicles interact, using varied control schemes for optimized performance.
They employ multi-agent reinforcement learning, model predictive control, and distributed optimization to minimize delays, enhance throughput, and improve safety.
Recent studies show measurable benefits including up to 23.3% travel time reduction, improved throughput by 9%, and significant reductions in collision rates.

A Mixed Traffic Control (MTC) Network is an urban or suburban traffic management system in which both human-driven vehicles (HVs) and robotic vehicles (RVs)—including autonomous vehicles (AVs) and connected automated vehicles (CAVs)—jointly traverse and interact in the road network. Control may be exerted at intersections via traffic signals, locally by RVs applying learned or rule-based policies, or even through distributed route guidance and real-time coordination. MTC frameworks aim to optimize objectives such as delay, throughput, safety, compliance, and energy in environments characterized by partial automation and heterogeneous agent behavior.

1. Network Architectures and Traffic Model Formulation

MTC networks are composed of nodes (intersections, junctions, stations) and edges (roads, lanes, ramps), with control points assigned to either fixed traffic signals, RVs/CAVs, or hybrid arrangements. Representative studies implement real-world subnetworks (e.g., a 14-intersection grid in Colorado Springs) featuring origin–destination (OD) demand patterns and vehicle diversity, with HVs modeled via continuous or discrete car-following dynamics (IDM, constant-acceleration models), and RVs switching to learned decision rules within specific control zones (commonly 30 m upstream of intersections) (Fan et al., 19 May 2025).

Key features include:

Multi-modal penetration: RV share can range from 0% (all-HV baseline) up to 100% with experimental configurations at intermediate rates (25%–80% RV).
Heterogeneous intersection control: Some intersections retain legacy fixed-time signalization, others are transitioned to direct RV or CAV control via Stop/Go policies or distributed feedback algorithms (Liu et al., 7 Apr 2025).
Traffic demand assignment: OD flows are represented by $F_{ij}$ (fraction from origin $i$ to destination $j$ ) and can be manipulated for sensitivity analysis, e.g., concentrating $p$ % of total flow on a subset of pairs (major corridors) or distributing evenly (Fan et al., 19 May 2025).

2. Control Algorithms: RL, MPC, and Hybrid Optimization

MTC networks deploy a spectrum of control methodologies, including:

Multi-Agent Reinforcement Learning (MARL):

Agents (RVs) operate in a decentralized partially-observable Markov decision process (POMDP) (Fan et al., 19 May 2025, Fan, 7 Dec 2025).
State space comprises local observations: queue lengths $q_d$ , average waiting times $\tau_d$ per direction $d$ , and intersection occupancy indicators $\sigma_d$ .
Action space is typically discrete (Stop, Go), but continuous controls (acceleration) are seen in bottleneck environments (Villarreal et al., 2023).
Policies and values are learned through variants of DQN (e.g., Rainbow DQN featuring C51 distributional output, prioritized replay) or PPO, with either separated or shared weights (Liu et al., 7 Apr 2025, Villarreal et al., 2023).
Control objective: minimizing average waiting time $\overline{W}$ , maximizing throughput, and explicitly penalizing unsafe (colliding) maneuvers.

Model Predictive Control (MPC) and Tube MPC:

Robust platoon control for strings of CAVs in mixed traffic environments adopts tube MPC, maintaining real-time trajectory deviations within precomputed invariant tubes under exogenous HDV uncertainty (Feng et al., 2019).
Event-triggered replanning drastically lowers computational and communication burden, only invoking centralized receding-horizon solves when state escapes the tube.
Quadratic cost criteria measure tracking error, control effort, and string-stability; explicit safety constraints (bumper–bumper, speed) enforce hard limits.

Distributed Optimization and Coordination:

At intersections, distributed penalization-enhanced block coordinate descent (Maximum Block Improvement algorithm) solves multi-agent mixed-integer quadratic programs (MIQP) partitioned per traffic light controller (TLC) and CAV (Le et al., 17 Sep 2024).
Local agents solve subproblems subject to coupling safety and signal constraints, with communication restricted to messages with neighbors only.
Person-by-person optimality and convergence are proven under mild feasibility conditions, with considerable reductions in computation time as compared to centralized MIQP.

Route Guidance & Flow Coordination:

Integrated MPC frameworks couple multi-class cell transmission models (CTM, for expressways) with macroscopic fundamental diagram (MFD) models for arterial regions, jointly optimizing route guidance, ramp metering, and perimeter flows (Di et al., 9 May 2024).
Cooperative guidance and control (CGC) schemes adjust OD splits, control rates, and metering based on predictive travel times, evaluated via trip completion flow and accumulations metrics.

3. Origin–Destination Patterns and Sensitivity

OD flow patterns are critical determinants of MTC network performance:

Concentrating flow on major corridors (e.g., NS+SN, NW+WN) with $p = 70\%$ or $p = 90\%$ fraction leads to distinct congestion profiles, which can be dynamically tuned to mitigate waiting time.
Uniform distribution of OD flows over all corridor pairs yields the lowest average network delay, indicating that spreading load across available capacity minimizes bottlenecks (e.g., $\overline{W} = 2.22$ s at 75% RV) (Fan et al., 19 May 2025).
Interaction effects: increasing RV penetration does not guarantee monotonic delay reduction; OD configuration and intersection control split strongly influence latent bottlenecks and critical queue formation.

4. Safety, Collision Rate, and Compliance

Safety is handled via explicit reward design (collision penalties), control-theoretic guarantees, and compliance schemes:

Collision rate (CR) is measured as $CR = N_\text{collided} / N_\text{departed}$ , empirically shown to decrease as more intersections are transitioned to RV control (minimum CR $=0.595\%$ at 80% RV in 12U+2S configuration) (Fan, 7 Dec 2025).
Removing conflicted left-turn movements generally lowers CR but can increase risk in certain network topologies.
Control barrier functions (CBFs) and Lyapunov-based quadratic programs for individual CAVs guarantee recursive feasibility and collision avoidance even under HDV uncertainty (Tzortzoglou et al., 17 Jun 2025).
Social compliance in route guidance is incentivized via refundable toll mechanisms, with the compliance probability adaptively learned and corrected via control Lyapunov functions—driving empirical compliance from initial 0 to $0.9$ within five decision points and reducing travel time by $23.3\%$ (Li et al., 28 Mar 2025).

5. Observation Spaces, Sensing, and Privacy

Mixed traffic control networks must reconcile scalable observation, privacy, and robustness:

Pixel-based (image) observation allows generic deployment across environments, without scenario-specific sensors or engineered state vectors; local BEV (bird’s-eye view) images (e.g., $84\times84$ ) encode both vehicle type and spatial context, enabling decentralized PPO policies to match or exceed the performance of precise-state methods (Villarreal et al., 2023).
Privacy-preserving crowdsensing reports only local RV penetration rates $P_v$ within a 30 m neighborhood, omitting trajectories and identities. Route selection is driven by scalar scores, reducing the shortage index by up to $69.4\%$ and mean waiting time by $27\%$ versus fixed signals, while keeping data strictly minimized (Wang et al., 2023).
Real-world hardware platforms instrument commercial vehicles via embedded compute, CAN-to-ROS bridges, and server-layer whitelisting, proving in-situ mixed autonomy traffic control feasibility for up to 100 vehicles over 22,752 miles (Nice et al., 2023).

6. Scalability, Limitations, and Future Directions

Tube MPC, distributed MIQP, and pixel-based RL scale effectively to networks of dozens of intersections and hundreds of vehicles, with per-agent computational and communicational load scaling sublinearly (Feng et al., 2019, Le et al., 17 Sep 2024, Villarreal et al., 2023).
Challenges remain in handling non-standard intersection geometries (roundabouts, ramps), partial penetration scenarios, real-world human driver stochasticity, and integrating macroscopic predictive models for urban-scale MTC networks.
Promising directions include hierarchical decomposition, asynchronous coordination, multi-objective reward extension (safety, emissions), formal privacy quantification (differential privacy $\varepsilon$ -budgets), and adaptive dynamic assignment of intersection control schemes (Fan et al., 19 May 2025, Di et al., 9 May 2024, Wang et al., 2023).

7. Summary of Empirical Performance

Network Feature	Best Observed Value	Reference
Minimum average waiting time	2.22 s @ 75% RV, uniform OD	(Fan et al., 19 May 2025)
Maximum throughput improvement	9% @ 80% RV, 8U+6S	(Liu et al., 7 Apr 2025)
Minimum collision rate	0.118% @ 60% RV, 12U+2S	(Fan, 7 Dec 2025)
Delay reduction via compliance	−23.3% travel time	(Li et al., 28 Mar 2025)
RV shortage index reduction	−69.4%	(Wang et al., 2023)
White phase delay reduction	3.2%–94.06% vs baseline	(Niroumand et al., 2022)
RC-H throughput (buses/private)	8000 veh/h (RC-H); signals < 4000 veh/h	(Chen et al., 2021)

In summation, Mixed Traffic Control Networks integrate multi-agent RL, robust optimization, cooperative and distributed control, and sensory-adaptive frameworks to provide scalable, safe, and efficient traffic management in urban environments characterized by heterogeneous vehicle autonomy and variable market penetration rates. Recent research demonstrates quantitative gains across throughput, delay, safety, and privacy metrics, establishing MTC as a viable paradigm for next-generation intelligent transportation systems.