Modern logistics companies face significant challenges in efficiently managing dynamic transportation systems, particularly in addressing Dynamic Vehicle Routing Problems (DVRP). These complex problems require specialized expertise for effective algorithm design. Traditional approaches often rely on manual algorithm design for specific routing problems. Reinforcement learning (RL)-based methods, though capable of learning heuristics for diverse routing problems, suffer from poor training efficiency and weak generalization to out-of-distribution (OOD) scenarios. To address these limitations, we leverage the strong generalization capabilities of large language models (LLMs) and adopt an LLM-based heuristic learning approach for DVRP. Our method learns heuristics and autonomously generates executable code within an evolutionary framework, requiring fewer than 10 samples for training. This enables fast training while maintaining robust performance on large-scale OOD instances. Comprehensive experiments across multiple routing problems—including Vehicle Routing Problems (VRP), VRP with Time Windows (VRPTW), DVRP, and DVRP with Time Windows (DVRPTW)—demonstrate that our approach learns effective heuristics and consistently surpasses Greedy baselines. Moreover, in constrained optimization tasks (VRPTW and DVRPTW), our method attains higher feasibility rates than Greedy baselines and approaches the performance of expert-designed heuristics. The proposed method presents a promising and scalable solution with significant potential for real-world industrial deployment.1

Deep Reinforcement Learning (DRL)-based frameworks, featuring Transformer-style policy networks, have demonstrated their efficacy across various Vehicle Routing Problem (VRP) variants. However, the application of these methods to the Multi-Trip Time-Dependent Vehicle Routing Problem (MTTDVRP) with maximum working hours constraints—a pivotal element of urban logistics—remains largely unexplored. This article introduces a DRL-based method called the Simultaneous Encoder and Dual Decoder Attention Model (SED2AM), tailored for the MTTDVRP with maximum working hours constraints. The proposed method introduces a temporal locality inductive bias to the encoding module of the policy networks, enabling it to effectively account for the time dependency in travel distance/time. The decoding module of SED2AM includes a vehicle selection decoder that selects a vehicle from the fleet, effectively associating trips with vehicles for functional multi-trip routing. Additionally, this decoding module is equipped with a trip construction decoder leveraged for constructing trips for the vehicles. This policy model is equipped with two classes of state representations, fleet state, and routing state, providing the information needed for effective route construction in the presence of maximum working hours constraints. Experimental results using real-world datasets from two major Canadian cities not only show that SED2AM outperforms the current state-of-the-art DRL-based and metaheuristic-based baselines but also demonstrate its generalizability to solve larger scale problems.

Dispatching large-scale fleets has been one of the fundamental aspects of managing heterogeneous truckload logistics. This operation involves optimization, visualization, and reporting of the resource plans meticulously crafted by expert planners and dispatchers on a daily basis. However, the limitations of human dispatchers, including errors in communication, routing, compliance, load planning, maintenance oversight, and neglect of driver preferences, can lead to lower customer satisfaction. We present FleetWiz, a Large Language Model-based (LLM) platform that enables logistics industry dispatchers to receive optimal recommendations based on real-time spatial data. FleetWiz seamlessly connects dispatchers, drivers, and resources, centralizing information within a unified resource-request network. This leads to enhanced transit times, reduced delays, and adaptive responses to dynamic conditions. It can execute tasks in different domains including filtering, optimizing, and answering questions based on the network of resources and requests. Specifically, a local Llama3 model is equipped with access to a geo-database for filtering, five different optimization methods for generating plans, and knowledge about the entire network and operations inside the company. Lastly, the tool's reliability, a generic interface for applying LLM agents alongside spatio-temporal optimization models, is demonstrated. The optimization models handle complex dispatching tasks requiring sequential reasoning, allowing the LLM to provide well-informed feedback based on the results.

The heterogeneous fleet and demand vehicle routing problem with time-window constraints (HFDVRPTW) is a crucial optimization problem of significant importance in real-world logistics operations. In this paper, we propose a deep reinforcement learning (DRL)-based method, termed spatial Edge-Feature EnhanCed mulTIgraph fusion encoder With spectral-based embedding and hieRarchical decOder with learnable TEmpoRal positional embedding (EFECTIW-ROTER, pronounced "Effective Router"), to tackle this complex and practical optimization problem. EFECTIW-ROTER utilizes two sparse graphs to represent node connectivity, where nodes correspond to customers and the depot. This sparsity results from the time-window constraints and customers' demand relative to the list of acceptable vehicle attributes specified for service within a heterogeneous fleet, determined by the reachability of the nodes based on these two factors. Leveraging two graph Transformer models, EFECTIW-ROTER's encoding module captures the interactions between the nodes based on these factors. One model encodes customers' heterogeneous demand with spatial edge features based on travel time between the nodes, while the second employs temporal positional embeddings to capture temporal relationships based on time-window ordering. A fusion model is introduced to integrate node interactions based on these graphs. Additionally, a spectral-attention-based pooling ensures effective state representation for the DRL-based method. EFECTIW-ROTER features a hierarchical attention decoder operating in two stages: heterogeneous vehicle selection and node selection. Enhanced with positional embeddings, the decoder is empowered to make effective routing decisions based on time-window constraints' ordering. Experimental results using real-world traffic data from two major Canadian cities confirm EFECTIW-ROTER's better performance over current state-of-the-art DRL-based and heuristic methods. EFECTIW-ROTER reduces travel times while also achieving faster computational times when compared to conventional heuristics. Additional experiments demonstrate its generalizability across larger instances.

Cascade ranking is widely used for large-scale top-k selection problems in online advertising and recommendation systems, and learning-to-rank is an important way to optimize the models in cascade ranking. Previous works on learning-to-rank usually focus on letting the model learn the complete order or top-k order, and adopt the corresponding rank metrics (e.g. OPA and NDCG@k) as optimization targets. However, these targets can not adapt to various cascade ranking scenarios with varying data complexities and model capabilities; and the existing metric-driven methods such as the Lambda framework can only optimize a rough upper bound of limited metrics, potentially resulting in sub-optimal and performance misalignment. To address these issues, we propose a novel perspective on optimizing cascade ranking systems by highlighting the adaptability of optimization targets to data complexities and model capabilities. Concretely, we employ multi-task learning to adaptively combine the optimization of relaxed and full targets, which refers to metrics Recall@m@k and OPA respectively. We also introduce permutation matrix to represent the rank metrics and employ differentiable sorting techniques to relax hard permutation matrix with controllable approximate error bound. This enables us to optimize both the relaxed and full targets directly and more appropriately. We named this method as Adaptive Neural Ranking Framework (abbreviated as ARF). Furthermore, we give a specific practice under ARF. We use the NeuralSort to obtain the relaxed permutation matrix and draw on the variant of the uncertainty weight method in multi-task learning to optimize the proposed losses jointly. Experiments on a total of 4 public and industrial benchmarks show the effectiveness and generalization of our method, and online experiment shows that our method has significant application value.

Efficient goods movement is a vital aspect of logistics and urban planning, impacting the flow of goods and the quality of life for residents. To aid in this, we present an interactive map-based system for visualizing and analyzing goods’ movements using GPS traces. The system takes raw GPS signal data and road network data as input, then performs preprocessing and spatial analysis on the data using Flask framework and Python scripts. The system offers a user-friendly interface to explore the patterns of goods movement dynamically. It displays the temporal and spatial movement trips in the city, providing an intuitive way to analyze and optimize goods movement. Our system’s ability to explore and visualize goods movement patterns makes it an essential addition to the existing literature on urban transportation analysis and a valuable tool for logistics companies and urban planners.

Truck speed prediction plays a key role in truck transportation management. However, it is a very challenging task since the truck traffic usually shows complex patterns. Most of the existing traffic prediction methods lack the ability to model the dynamic spatial-temporal correlations of truck traffic or ignore contributing contextual factors that impact traffic. Also, truck traffic data is typically sparse and noisy, which makes the truck speed prediction an even more challenging task. How to improve the truck speed prediction by taking advantage of other relevant truck traffic information (such as the truck flow) has not been investigated in depth. Additionally, traffic congestions and poor driving conditions caused by extreme weather conditions can make sudden changes in the general pattern of the truck speed. In this paper, we propose a novel Multi-Task Context Based Gated Recurrent Unit Graph Convolutional Network (MT-C2G) to predict the truck speed under extreme weather conditions. MT-C2G includes four major components: The spatial dependence learning component captures the spatial dependencies shaped by the topological structure of the road network. Truck traffic feature temporal dependence modeling component is built to acquire the temporal dependencies involved in the truck traffic features, and contextual feature temporal dependence modeling component employs a layer of GRU units to capture the temporal dependencies of contextual factors. The multi-task learning component then leverages the information between the truck speed and flow prediction tasks through attention mechanism for improving the performance. Moreover, a data augmentation method SMOTE is utilized to balance the data with the extreme weather conditions. Experiments on two real datasets demonstrate that the proposed MT-C2G fairly outperforms six state-of-the-art traffic prediction methods.

Most machine learning classifiers only concern classification accuracy, while certain applications (such as medical diagnosis, meteorological forecasting, and computation advertising) require the model to predict the true probability, known as a calibrated estimate. In previous work, researchers have developed several calibration methods to post-process the outputs of a predictor to obtain calibrated values, such as binning and scaling methods. Compared with scaling, binning methods are shown to have distribution-free theoretical guarantees, which motivates us to prefer binning methods for calibration. However, we notice that existing binning methods have several drawbacks: (a) the binning scheme only considers the original prediction values, thus limiting the calibration performance; and (b) the binning approach is non-individual, mapping multiple samples in a bin to the same value, and thus is not suitable for order-sensitive applications. In this paper, we propose a feature-aware binning framework, called Multiple Boosting Calibration Trees (MBCT), along with a multi-view calibration loss to tackle the above issues. Our MBCT optimizes the binning scheme by the tree structures of features, and adopts a linear function in a tree node to achieve individual calibration. Our MBCT is non-monotonic, and has the potential to improve order accuracy, due to its learnable binning scheme and the individual calibration. We conduct comprehensive experiments on three datasets in different fields. Results show that our method outperforms all competing models in terms of both calibration error and order accuracy. We also conduct simulation experiments, justifying that the proposed multi-view calibration loss is a better metric in modeling calibration error. In addition, our approach is deployed in a real-world online advertising platform; an A/B test over two weeks further demonstrates the effectiveness and great business value of our approach.