Dual-Population Algorithm Accelerates EV Routing Solutions
In the rapidly evolving landscape of green logistics, one of the most persistent challenges has been optimizing the operational efficiency of electric vehicle (EV) fleets. Unlike traditional delivery models, EV routing is complicated by dual constraints—battery capacity and cargo load—that must be balanced in real time. This complex problem, known as the Electric Vehicle Routing Problem (EVRP), has long posed computational hurdles for logistics planners aiming to minimize both energy consumption and delivery costs. Now, a breakthrough algorithm developed at Anhui University promises to redefine how companies approach EV fleet management, offering faster convergence and higher-quality solutions than previously possible.
The new method, introduced in a study published in CAAI Transactions on Intelligent Systems, leverages a novel dual-population co-evolutionary framework to tackle the intricacies of EVRP with unprecedented efficiency. Led by Dr. Chao Wang, an associate professor in the School of Artificial Intelligence at Anhui University, the research team has engineered a system that not only accelerates solution discovery but also enhances the robustness of route planning under dynamic real-world conditions.
At its core, the algorithm addresses a fundamental limitation in existing optimization techniques: the overwhelming search space created when simultaneously managing vehicle paths and charging schedules. Traditional approaches—ranging from exact methods like branch-and-bound to heuristic strategies such as simulated annealing or tabu search—often struggle to scale effectively. Exact methods become computationally prohibitive as problem size increases, while heuristic algorithms are prone to getting trapped in local optima, especially when initial solutions are suboptimal or neighborhood operators are poorly chosen.
Evolutionary algorithms have offered some relief, using population-based parallel search to explore broader solution landscapes. However, even these methods face difficulties when applied to EVRP due to the tight coupling between routing decisions and charging logistics. Prior attempts using differential evolution or genetic algorithms have shown limited success, either failing to produce high-quality feasible solutions or suffering from slow convergence.
The innovation proposed by Wang and his colleagues lies in reframing the problem through a two-tiered evolutionary process. Instead of attacking the full EVRP directly, the team introduces a simplified auxiliary problem—essentially a Capacitated Vehicle Routing Problem (CVRP) without electricity constraints. By treating charging stations as passive waypoints with no service demand, they create a parallel optimization task that is inherently easier to solve. This simpler CVRP variant converges more quickly to feasible routes, providing a rich source of high-quality path structures that can inform the more complex EVRP.
What sets this approach apart is not just the use of a secondary problem, but the intelligent mechanism for knowledge transfer between the two domains. The researchers recognized that CVRP and EVRP, while related, are structurally different—making direct solution exchange ineffective. To bridge this gap, they designed a feature representation method based on an enhanced distance adjacency matrix. This matrix goes beyond simple Euclidean distances by incorporating relative proximity metrics that reflect vehicle assignment patterns within routes. In doing so, it encodes both spatial relationships and fleet allocation logic into a unified format, enabling machine learning models to interpret and transform solutions across problem types.
Central to the system’s success is the integration of a denoising autoencoder (DAE), a neural network architecture commonly used in transfer learning. The DAE is trained to map solution representations from the CVRP domain to those in the EVRP domain, effectively learning the transformation rules needed to adapt a fuel-agnostic route into one that respects battery limitations. During the evolutionary process, elite solutions from each population are periodically converted using the trained DAE and injected into the other population as offspring. This bidirectional migration allows the CVRP population to benefit from realistic energy-aware routing patterns, while the EVRP population gains access to well-structured, capacity-feasible routes generated under less constrained conditions.
This co-evolutionary dynamic creates a feedback loop where both populations improve in tandem. The CVRP side benefits from exposure to energy-conscious routing strategies, gradually evolving solutions that are not only efficient in terms of distance and load but also compatible with EV operational realities. Meanwhile, the EVRP population receives a steady influx of high-quality route skeletons, significantly reducing the time required to discover viable charging sequences. The result is a synergistic optimization process that outperforms conventional single-population methods in both speed and solution quality.
To validate their approach, the team conducted extensive experiments using standardized EVRP benchmark datasets, including medium-scale instances with 200 customers and large-scale problems involving 400 delivery points. These datasets, widely used in academic research, simulate real-world urban delivery scenarios with realistic constraints on vehicle capacity, battery life, and charging infrastructure availability.
The proposed Dual-Population Co-Evolutionary Algorithm (COEA) was benchmarked against five state-of-the-art methods: BACO, KBEA, HVNS, ALNS, and TS-MCWS. These include both heuristic techniques and advanced evolutionary algorithms, ensuring a rigorous comparison across multiple paradigms. All algorithms were configured with identical population sizes and maximum evaluation counts to ensure fairness, with each test case run independently 20 times to account for stochastic variability.
Performance was assessed using two primary metrics: total travel distance and fleet size, combined into a composite cost function weighted to reflect the relative importance of energy efficiency and capital utilization. Additionally, for multi-objective variants, the hypervolume (HV) indicator was used to evaluate the quality of the Pareto front, capturing both convergence and diversity of non-dominated solutions.
The results were compelling. Across 18 test instances, COEA achieved the best-known travel distance in 11 cases, consistently outperforming all comparison algorithms. More importantly, it did so while maintaining or reducing the number of vehicles required—a critical factor in commercial logistics where fleet size directly impacts operational costs. In several large-scale scenarios, COEA reduced total mileage by over 15% compared to the next best performer, demonstrating its scalability and robustness.
Perhaps even more telling was the algorithm’s convergence behavior. When tracking average solution cost over successive generations, COEA exhibited a significantly steeper improvement curve than its competitors. Within the first 20% of the total iteration budget, it had already surpassed the final performance of some baseline methods. This rapid convergence is particularly valuable in practical applications where decision-makers need timely recommendations rather than waiting for prolonged optimization runs.
Further analysis revealed that the key components of COEA—namely the improved distance adjacency matrix and the DAE-based knowledge transfer—each contributed meaningfully to the overall performance. Ablation studies showed that removing either component led to measurable degradation in solution quality. When solutions were exchanged without the DAE transformation (Variant II), the lack of semantic alignment between problem domains resulted in inefficient migrations and slower progress. Similarly, when the adjacency matrix was not enhanced with vehicle assignment information (Variant I), the model struggled to capture the structural nuances of multi-vehicle routing, leading to suboptimal clustering and path formation.
Compared to a standalone MOEA/D implementation on the original EVRP, the full COEA framework demonstrated superior performance across all metrics. This underscores the strategic value of leveraging auxiliary problems in complex optimization tasks. The study suggests that the principle of co-evolution between simplified and full-fidelity models could be generalized to other domains where constraint interactions create computational bottlenecks.
From an industry perspective, the implications of this research are substantial. As global logistics networks face mounting pressure to decarbonize, electric vehicles are no longer a niche alternative but a core component of sustainable supply chains. Companies like Amazon, UPS, and DHL have already committed to electrifying large portions of their delivery fleets, driven by regulatory mandates and consumer expectations. However, the transition is not without operational risks. Poorly optimized EV routes can lead to unexpected battery depletion, missed deliveries, and inefficient use of charging infrastructure—all of which erode the economic and environmental benefits of electrification.
By delivering faster and more reliable route plans, COEA enables logistics operators to maximize vehicle uptime, reduce energy waste, and maintain service levels even in dense urban environments. The ability to quickly generate high-quality solutions also supports dynamic re-routing in response to traffic disruptions, weather events, or last-minute order changes—capabilities that are increasingly essential in today’s on-demand economy.
Moreover, the algorithm’s design reflects a growing trend in artificial intelligence: the integration of machine learning with classical optimization techniques. Rather than replacing traditional operations research methods, the DAE acts as a smart translator, enhancing the interoperability of different problem formulations. This hybrid approach combines the interpretability and constraint-handling strength of evolutionary algorithms with the pattern recognition power of deep learning, creating a system greater than the sum of its parts.
The research also highlights the importance of problem representation in algorithmic design. The improved distance adjacency matrix serves as a bridge between symbolic AI and data-driven modeling, allowing geometric and topological features of routing problems to be captured in a format amenable to neural processing. This kind of feature engineering—thoughtfully designed rather than automatically learned—ensures that domain knowledge remains central to the optimization process, aligning with best practices in trustworthy AI development.
Looking ahead, the team at Anhui University sees opportunities to extend their framework to more complex variants of EVRP, including time windows, heterogeneous fleets, and partial recharging capabilities. Future work may also explore adaptive training schedules for the DAE, online learning from real-world deployment data, and integration with digital twin systems for predictive logistics planning.
For logistics technology providers, the findings offer a blueprint for next-generation routing engines. Rather than relying solely on incremental improvements to existing solvers, developers can now consider architectures that exploit cross-domain knowledge transfer and co-evolutionary dynamics. The success of COEA suggests that some of the most significant gains in optimization performance may come not from brute-force computation, but from smarter ways of structuring the search process itself.
In conclusion, the dual-population co-evolutionary algorithm represents a significant advancement in the field of green logistics optimization. By intelligently decomposing the EVRP into interrelated subproblems and enabling knowledge flow between them, it achieves a level of efficiency and solution quality that sets a new standard for the industry. As electric vehicles continue to reshape the transportation landscape, innovations like this will play a crucial role in ensuring that the promise of sustainable logistics becomes a practical reality.
Dual-Population Algorithm Accelerates EV Routing Solutions
Chao Wang, Fang Qin, Rongrong Liu, Hao Jiang, School of Artificial Intelligence, Anhui University
CAAI Transactions on Intelligent Systems, DOI: 10.11992/tis.202209007