🤖 machine learning

Simulating Cognitive Smart Freight Corridors with Agent-Based Models and Reinforcement Learning

This paper proposes an agent-based modeling framework integrating reinforcement learning to simulate cognitive smart freight corridors, demonstrating that adaptive multi-agent coordination significantly improves throughput, reduces congestion, and optimizes energy usage compared to traditional rule-based approaches.

Original authors: Madelaine Martinez-Ferguson, Chun Wang, Mustafa Can Camur, Xueping Li

Published 2026-08-27

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Madelaine Martinez-Ferguson, Chun Wang, Mustafa Can Camur, Xueping Li

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The movement of goods across a country relies on a vast network of roads, where millions of trucks carry the materials that build our cities and fill our stores. For decades, the efficiency of this system has been threatened by traffic jams, which waste time, fuel, and money. In recent years, engineers have looked to connected and automated vehicles—trucks that can talk to each other and to the road infrastructure—to solve these problems. The idea is that if trucks can communicate instantly, they can travel closer together in tight groups, known as platoons, which reduces air resistance and saves energy. However, simply putting smart trucks on a highway is not enough. The real challenge lies in how these vehicles interact with the road itself, how they decide when to stop for charging, and how they react when accidents or sudden surges in traffic occur. Testing these complex interactions in the real world is dangerous and prohibitively expensive, so researchers must turn to sophisticated computer simulations to understand what might happen before any physical changes are made.

A team of researchers has developed a new way to simulate these future freight corridors, treating the road, the communication network, and the decision-making software as a single, living system. Instead of relying on fixed rules that tell trucks exactly what to do, they created a digital environment where the trucks and the road controllers learn from experience. This approach uses a method called reinforcement learning, where a computer program tries different actions, sees what happens, and gradually improves its strategy to keep traffic flowing smoothly. The researchers built a virtual highway with twenty segments, five charging stations, and four terminals where trucks begin and end their journeys. They populated this digital world with 1,500 trucks, some driven by humans and others operating autonomously, and watched how they behaved under different levels of technology.

The study compared three distinct scenarios to see how much difference the technology makes. The first was a baseline scenario representing today's reality, where all trucks are driven manually, there is no communication between vehicles, and drivers make their own decisions about when to charge. The second was an assisted scenario, where trucks could talk to the road and each other, allowing them to see traffic conditions and charging availability, but they still followed simple, pre-programmed rules for forming groups and stopping. The third was a cognitive scenario, the most advanced version, where the trucks were fully autonomous and the road itself was managed by an intelligent system that learned to optimize traffic flow and charging in real time. In this advanced setup, a central controller managed the formation of truck groups and the opening of special lanes, while separate agents at each charging station learned to adjust prices and priorities to keep lines moving.

The results of these simulations revealed a clear hierarchy of performance. The assisted scenario, which added communication but kept simple rules, showed meaningful improvements in energy efficiency. By allowing trucks to travel in closer groups, the system reduced the energy needed to drive each kilometer by about 7.5 percent. However, the cognitive scenario, where the road and trucks learned to work together, delivered far more dramatic benefits. This intelligent system increased the number of trips completed per hour by 27 percent compared to the baseline and reduced the measure of traffic congestion by nearly 64 percent. The key difference was that the cognitive system did not just react to traffic; it actively managed the flow, keeping trucks moving even when the road was heavily used. While the total amount of carbon dioxide emitted was higher in the cognitive scenario, this was simply because so many more trucks were successfully completing their journeys; the emissions per mile driven remained low and comparable to the other scenarios.

The researchers also tested how well these systems held up when things went wrong. They simulated accidents that blocked parts of the road, severe weather that slowed traffic, charging station failures, and sudden spikes in the number of trucks trying to enter the highway. In these stressful conditions, the cognitive system proved remarkably resilient. When accidents occurred, the assisted system collapsed, with travel times spiking by 155 percent because its rigid rules could not adapt to the blockage. In contrast, the cognitive system maintained near-normal performance, adjusting its strategies on the fly to keep traffic moving. It also handled sudden increases in demand better than the other systems, actually completing more trips than usual when the volume of trucks doubled. The only time the assisted system performed slightly better was when a charging station failed, as its simpler rules allowed trucks to spread out naturally without needing a central coordinator to reorganize them.

These findings suggest that the future of freight transport will likely depend on more than just smarter trucks. While connecting vehicles and giving them basic automation offers immediate gains in fuel savings, the greatest improvements in speed and reliability come from an intelligent infrastructure that can learn and adapt. The study indicates that as demand for freight grows, the advantage of having a cognitive corridor widens, making it a critical investment for high-traffic routes. The researchers note that their work is based on simulations, and while the results are promising, real-world testing will be needed to confirm these outcomes. They also point out that their model currently focuses on a one-way highway and could be expanded to include more complex road networks and dynamic pricing strategies. Nevertheless, the simulation provides a clear roadmap: by treating the road, the vehicles, and the data as a single learning system, we can build freight corridors that are not only faster and cleaner but also far more robust against the inevitable disruptions of daily life.

Technical Summary: Simulating Cognitive Smart Freight Corridors with Agent-Based Models and Reinforcement Learning

Problem Statement

The U.S. trucking industry faces severe cost inefficiencies due to traffic congestion, which cost the sector $108.8 billion in lost productivity in 2022 alone. While Connected and Automated Vehicles (CAVs) offer potential solutions through cooperative platooning and vehicle-to-everything (V2X) communication, realizing these benefits requires more than vehicle-level technology. The magnitude of benefits depends nonlinearly on penetration rates, infrastructure design, and control policies.

Existing simulation approaches suffer from critical gaps:

Lack of Joint Adaptation: Most models address individual aspects (e.g., RL for platooning or charging control) but fail to capture joint adaptive behaviors across physical, cyber, and decision layers.
Static Policies: Current decision layers rely on predefined optimization models or fixed control rules that cannot adapt to stochastic freight demand, disruptions, or evolving conditions.
Testing Limitations: Physical experimentation is expensive, and real-world pilots cannot cover the full combinatorial space of infrastructure configurations and control policies.

There is a need for a framework that integrates heterogeneous freight agents, treats V2X-enabled corridor infrastructure as an active control layer, and learns joint policies for platoon formation, charging coordination, and disruption response.

Methodology

The authors propose a modular, Python-based Agent-Based Modeling (ABM) framework structured as a three-layer system:

1. Three-Layer Architecture

Physical Layer: Comprises road segments (modeled with a modified Bureau of Public Roads function using $\beta=2$ for freight realism), managed lanes, and charging stations (150–350 kW). Freight terminals act as origin-destination nodes.
Connectivity (Cyber) Layer: Enables V2X communication (V2V and V2I) with configurable penetration rates and fixed 50ms latency. This layer separates infrastructure capability from control policy, allowing for scenario testing.
Decision Layer: Implements decentralized control via two agent types:
- Corridor Controller (RL): A single Deep Q-Network (DQN) agent managing platoon formation and lane activation. It observes a 14-dimensional state vector (congestion, utilization, queue length, etc.) and selects from five discrete actions (e.g., encourage platoons, activate managed lanes).
- Station Agents (MARL): Five independent Multi-Agent Reinforcement Learning agents managing local pricing and priority. They communicate over two rounds per step to coordinate load distribution, selecting from five pricing/priority actions.

2. Simulation Mechanics

Agents: Freight trucks maintain state vectors including location, State of Charge (SoC), autonomy mode (manual, assisted, autonomous), and platoon membership.
Energy Model: Solo driving consumes 1.2 kWh/km; platooning followers achieve a 15% reduction ( $\eta_p = 0.15$ ).
Training: The system uses a multi-objective reward function balancing throughput, delay, energy, and instability. Training involves 200 episodes (10 simulated hours each) with experience replay and target network updates.

3. Experimental Scenarios

Three scenarios were evaluated on the same physical infrastructure (20 segments, 5 stations, 1,500 trucks):

Baseline: 100% manual trucks, no V2X, fixed routes, greedy charging.
Assisted: 70% assisted (SAE L2/L3) trucks with full V2X. Uses rule-based platooning and greedy charging assignment.
Cognitive: 70% autonomous (SAE L4/L5) trucks with full V2X. Employs trained DQN for corridor control and MARL for dynamic pricing/priority.

Key Contributions

Integrated Framework: Development of a modular ABM environment coupling physical infrastructure, V2X connectivity, and adaptive decision-making layers.
Hybrid RL/MARL Control: Integration of single-agent RL for global corridor management (platoons/lanes) and distributed MARL for local charging coordination, a combination not previously applied to freight corridors.
Comprehensive Evaluation: Quantitative assessment across throughput, congestion, energy, emissions, and robustness metrics.
Zero-Shot Robustness: Demonstration of policy generalization to disruption scenarios (accidents, weather, station failures, demand spikes) without retraining.

Preliminary Results

Experiments were conducted over 10-hour horizons with five replications per scenario.

Performance Metrics

Throughput: The Cognitive scenario achieved 27.0% higher throughput (98.27 trips/hr) compared to the Baseline (77.38 trips/hr).
Congestion: The Cognitive scenario reduced the congestion index by 63.9% (0.15 vs. 0.42) relative to the Baseline.
Energy & Emissions: Both Assisted and Cognitive scenarios reduced energy consumption per kilometer by approximately 7.5–7.9% due to platooning. Total CO2 emissions were higher in the Cognitive scenario solely due to the increased volume of completed trips, though per-kilometer emissions remained comparable.
Travel Time: Cognitive reduced average travel time by 7.2% (1.33 hr vs. 1.43 hr).

Sensitivity Analysis

Demand: Under high demand (multiplier $\ge$ 1.5x), the throughput and congestion advantages of the Cognitive scenario widened significantly, whereas rule-based policies (Assisted) saturated.
Charging Infrastructure: Cognitive throughput remained stable (98–101 trips/hr) across varying port counts (5–25), while Baseline and Assisted showed lower throughput at low port counts. MARL coordination demonstrated diminishing returns beyond 15 ports but maintained superior performance over rule-based assignment.

Disruption Robustness

In a zero-shot evaluation (no retraining):

Cognitive maintained near-full performance (Robustness score $\approx$ 1.00) under accidents, weather, and demand spikes. Notably, it exceeded baseline throughput during demand spikes (1.31).
Station Failure: In this specific scenario, Assisted marginally outperformed both Cognitive and Baseline (Robustness scores: Assisted 1.02 vs. Cognitive 0.99 and Baseline 0.99) by redistributing charging demand across available ports without centralized coordination.
Accidents: Assisted collapsed under accident conditions (Robustness 0.00) due to a 155% travel time increase caused by platoon congestion at bottlenecks, highlighting the vulnerability of rule-based policies to capacity shocks.
Baseline showed moderate resilience but lower overall throughput.

Computational Efficiency

Inference latency was benchmarked at 47 $\mu$ s per simulation step (3.8 $\mu$ s for DQN + 42.9 $\mu$ s for MARL communication), four orders of magnitude below the 5-minute control interval, confirming feasibility for real-time deployment on edge hardware.

Significance and Claims

The paper claims that this work addresses a critical gap in assessing smart corridor operations by providing the first framework to simultaneously integrate ABM for heterogeneous agents, V2X as an active control layer, and adaptive learning for joint policy optimization.

The authors conclude that:

V2X and Assisted Automation deliver measurable near-term efficiency gains, particularly in energy consumption per kilometer.
RL/MARL Control offers superior throughput scaling and congestion relief, specifically under high-demand conditions where static rules fail.
Resilience: Cognitive corridors exhibit superior robustness to most disruption types compared to rule-based systems, though rule-based systems may show marginal advantages in specific station failure scenarios. This suggests that future corridor management should evolve from static coordination toward adaptive, intelligent infrastructure.

The authors maintain a modest tone, noting that results are "directional and preliminary" due to stochastic variations in corridor generation and the limited number of replications. They identify future work needs, including expanding to bidirectional networks, dynamic rerouting, and validation against real-world freight data.