Dual-Population Algorithm Speeds Up EV Routing Solutions
In the rapidly evolving landscape of green logistics, a new computational breakthrough is poised to transform how electric vehicle (EV) fleets are managed. Researchers from the School of Artificial Intelligence at Anhui University have developed an innovative algorithm that significantly accelerates the solution of complex electric vehicle routing problems (EVRP), addressing one of the most pressing challenges in sustainable urban transportation and last-mile delivery.
The study, led by Wang Chao, Qin Fang, Liu Rongrong, and Jiang Hao, introduces a dual-population co-evolutionary algorithm (COEA) designed to optimize both route planning and charging strategies simultaneously—a task that has long plagued logistics operators due to its computational complexity. As cities worldwide push for zero-emission transport systems, the ability to efficiently manage EV fleets is no longer just an operational concern but a strategic imperative for companies aiming to meet environmental targets while maintaining cost efficiency.
Traditional vehicle routing problems (VRP) have been studied extensively for decades, with algorithms capable of handling constraints such as vehicle capacity, time windows, and customer demand. However, the introduction of electric vehicles adds a critical new dimension: energy management. Unlike internal combustion engine vehicles, EVs require careful planning around battery capacity, charging infrastructure availability, and recharging time. This transforms the classic capacitated vehicle routing problem (CVRP) into a far more complex variant—EVRP—where solutions must satisfy both cargo load limits and energy constraints.
Existing methods for solving EVRP fall into two broad categories: exact algorithms and heuristic approaches. While exact methods like branch-and-bound or mixed-integer programming can guarantee optimal solutions, they become computationally infeasible as problem size increases. For real-world applications involving hundreds of customers and multiple depots, these techniques often fail to deliver timely results. On the other hand, heuristic and metaheuristic algorithms such as simulated annealing, tabu search, and adaptive large neighborhood search (ALNS) offer faster computation but are prone to getting stuck in local optima, especially when dealing with the coupled nature of routing and charging decisions.
Evolutionary algorithms have shown promise in tackling such multi-constraint optimization problems due to their population-based search mechanism, which explores diverse regions of the solution space in parallel. However, even advanced genetic algorithms struggle with the dual objectives of minimizing total travel distance and reducing fleet size while ensuring all vehicles remain within battery limits throughout their routes.
This is where the innovation from Anhui University stands out. The team’s COEA framework does not attempt to solve the full EVRP directly. Instead, it employs a clever divide-and-conquer strategy by creating a simplified auxiliary problem—essentially a CVRP without electricity constraints—while retaining the same customer distribution and depot configuration. In this simplified model, charging stations are treated as passive nodes, allowing the algorithm to focus solely on optimizing delivery sequences and vehicle assignments.
What makes this approach revolutionary is not just the use of a secondary problem, but how information flows between the two domains. The researchers recognized that simply transferring raw route data from the simple CVRP to the complex EVRP would be ineffective due to the structural differences between the two problems. To bridge this gap, they introduced a novel solution representation method based on an enhanced distance adjacency matrix.
This matrix goes beyond conventional distance tables by embedding additional contextual information about vehicle assignments and route segmentation. Each entry in the matrix reflects not only the physical distance between two points but also a weighted value that indicates whether those points are likely to be served by the same vehicle. This allows the algorithm to implicitly capture clustering patterns—groups of customers that should be grouped together on a single route—making the encoded solutions richer in semantic meaning.
But the real breakthrough lies in the mechanism used to translate knowledge across problem types. Drawing inspiration from machine learning, particularly transfer learning, the team implemented a denoising autoencoder (DAE)—a type of neural network commonly used for feature extraction and data reconstruction. The DAE is trained to map solution structures from the CVRP domain to their corresponding counterparts in the EVRP space, effectively learning the transformation rules that account for energy consumption, charging stops, and battery dynamics.
The process works iteratively. During the evolutionary cycle, elite solutions from both populations—those showing the best trade-offs between route length and vehicle count—are selected and converted into their matrix representations. These matrices serve as input-output pairs to train the DAE, refining its ability to predict how an optimal CVRP route might need to be adjusted to accommodate charging requirements in the EVRP context.
Once trained, the DAE acts as a knowledge transfer engine. It takes high-quality CVRP solutions and generates plausible EVRP equivalents by inserting charging stops at strategic locations, guided by battery depletion models and road network data. Conversely, good EVRP solutions are fed back into the CVRP population, helping to shape the auxiliary problem toward configurations that are more likely to yield viable electric routes. This bidirectional exchange creates a synergistic loop, where each population benefits from the other’s progress.
A key advantage of this architecture is convergence speed. Because the CVRP subproblem lacks the complicating factor of energy constraints, it reaches feasible, high-quality solutions much faster than the full EVRP. By periodically injecting these advanced CVRP solutions—transformed via the DAE—into the main EVRP population, the algorithm jump-starts the search process, avoiding prolonged exploration of infeasible regions.
The researchers tested their COEA against five state-of-the-art algorithms: BACO, KBEA, HVNS, ALNS, and TS-MCWS. They used a standard EVRP benchmark set containing 18 instances, ranging from medium-scale (200 customers) to large-scale (400 customers). All algorithms were run under identical conditions—same population size, maximum evaluations, and independent trials—to ensure fair comparison.
Results showed that COEA consistently outperformed the competition in terms of solution quality and convergence rate. On average, it achieved shorter total travel distances while using fewer vehicles across most test cases. In 11 out of 18 instances, COEA found the best-known solutions, particularly excelling in large-scale scenarios where computational efficiency matters most. When compared to BACO, HVNS, and TS-MCWS, the improvements were substantial, with reductions in total cost metrics exceeding 20% in several cases.
Even more telling was the performance relative to KBEA, another evolutionary algorithm known for leveraging historical search patterns to guide crossover and mutation operations. While KBEA performed well—especially in large problems, where it secured the best result in seven instances—COEA still managed to surpass it in the majority of tests. This suggests that the knowledge transfer enabled by the dual-population design provides a more effective search bias than relying solely on intra-population learning.
To further validate the robustness of their approach, the team conducted ablation studies by disabling key components of the COEA framework. One variant removed the improved distance adjacency matrix, reverting to basic sequence encoding. Another eliminated the DAE-based knowledge transfer, replacing it with direct solution exchange between populations. A third baseline used a standard MOEA/D algorithm to solve EVRP without any auxiliary problem.
The results were clear: each component contributed significantly to overall performance. Removing the enhanced matrix representation led to slower convergence and lower solution quality, confirming that richer solution encoding enables more effective learning. Disabling the DAE caused the algorithm to lose its adaptive transfer capability, resulting in solutions that were less feasible and more scattered in the objective space. And compared to the standalone MOEA/D, the full COEA demonstrated superior exploration-exploitation balance, avoiding premature convergence.
Perhaps the most compelling evidence came from the hypervolume (HV) metric, which measures both the diversity and convergence of a multi-objective solution set. Across large-scale test problems, COEA achieved higher HV values than all competing algorithms, indicating that it not only finds better solutions but maintains a broader front of trade-off options between minimizing distance and reducing fleet size.
From a practical standpoint, the implications of this research are significant. Logistics companies operating EV fleets can leverage such algorithms to reduce operational costs, extend vehicle range through smarter routing, and improve service reliability. Municipalities planning urban delivery zones or curbside charging networks can use these models to simulate fleet behavior and optimize infrastructure placement.
Moreover, the methodology opens up new possibilities for cross-domain optimization in transportation. The concept of using a simplified proxy problem to guide the solution of a more complex one could be applied to other challenging domains, such as drone delivery routing with battery and weather constraints, multimodal urban mobility planning, or even dynamic ride-sharing systems with mixed vehicle types.
The success of COEA also highlights a growing trend in computational logistics: the fusion of classical operations research with modern machine learning techniques. Rather than treating AI as a black box, the Anhui University team integrated deep learning as a structured component within an evolutionary framework, using it to enhance—not replace—traditional optimization principles. This hybrid approach respects the mathematical rigor of VRP modeling while harnessing the pattern recognition power of neural networks.
As electric mobility continues to scale, the demand for intelligent routing systems will only grow. Current EV routing tools often rely on rule-based heuristics or simplified assumptions about charging behavior, leading to suboptimal plans in real-world conditions. The COEA algorithm represents a step forward in building truly adaptive, data-driven decision support systems that can handle the full complexity of electrified logistics.
The research also underscores the importance of algorithmic innovation in achieving sustainability goals. While much attention is given to battery technology and charging hardware, software intelligence plays an equally crucial role in maximizing the environmental and economic benefits of EV adoption. Efficient routing reduces energy consumption, extends battery life, and minimizes the need for oversized fleets—all of which contribute to lower carbon footprints.
Looking ahead, the authors suggest several directions for future work. These include incorporating real-time traffic data, integrating partial recharging strategies, and extending the model to handle heterogeneous fleets with different battery capacities and charging speeds. Additionally, the team is exploring ways to make the DAE training process more efficient, potentially using online learning to adapt to changing network conditions.
In conclusion, the dual-population co-evolutionary algorithm developed by Wang Chao, Qin Fang, Liu Rongrong, and Jiang Hao at Anhui University marks a significant advancement in the field of green logistics. By combining evolutionary computation with transfer learning, the method offers a powerful new tool for solving some of the most challenging problems in electric vehicle fleet management. As cities and corporations accelerate their transition to sustainable transport, innovations like COEA will be essential in turning ambitious climate goals into practical, scalable solutions.
Dual-Population Algorithm Speeds Up EV Routing Solutions
Wang Chao, Qin Fang, Liu Rongrong, Jiang Hao, School of Artificial Intelligence, Anhui University, CAAI Transactions on Intelligent Systems, DOI: 10.11992/tis.202209007