Publication

The real-world delivery route optimization problem aims to design routes that can adapt to complex delivery scenarios by learning and understanding drivers’ behaviors. However, the complexity of real-world environment makes it difficult to quantify and incorporate various factors influencing delivery. To overcome this challenge, this paper proposes a high-fidelity approximation of real-world delivery routes. In particular, we develop a two-stage architecture, where in the first stage, we leverage a pure machine learning model to learn drivers’ delivery behaviors, while in the second stage, a classical TSP model is embedded for respecting the shortest-path behavior locally. To realize the representation, learning and inference of drivers’ delivery behaviors, we build our approach on a language model in machine learning. Specifically, a real-world delivery route is first analogously defined as a sentence in natural language, which preserves the implicit knowledge of drivers’ behavioral patterns. Then, we use an unsupervised learning-based method to learn the vector representations of delivery routes and infer the drivers’ delivery sequences based on a vector-to-sequence algorithm. Numerical results show that the proposed approach outperforms a pure optimization approach and is competent to that of a well-designed tailored non-machine learning method. Moreover, the proposed architecture has its generalizability to variants of the problem. First, as a versatile method, the proposed model can easily be extended to different data sources from different delivery service providers. Second, for real-world shortest-path problems, the proposed model can also be used to efficiently provide routes that allow for various realistic factors such as safety and convenience by learning from the drivers’ experiences.

The vehicle dispatching system is one of the most critical problems in online taxi-hailing platforms, which requires adapting the operation and management strategy to the dynamics of demand and supply. In this paper, we propose a single-agent deep reinforcement learning approach for the vehicle repositioning problem by reallocating vacant vehicles to regions with a large demand gap in advance. The simulator and the vehicle repositioning algorithm are designed based on industrial-scale real-world data and the workflow of online taxi-hailing platforms, ensuring the practical value of our approach. Besides, the vehicle repositioning problem is translated in analogy with the load balancing problem in computers. Inspired by the recommendation system, the high concurrency of repositioning requests is addressed by sorting the actions as a recommendation list, whereby matching action with requests. Experiments demonstrate that the proposed approach is superior to the existing ones. It is also worth noting that the proposed approach won first place in the vehicle repositioning task of KDD Cup 2020.

Species identification is one of the key steps in the management and conservation planning of many forest ecosystems. We introduce Deep BarkID, a portable tree identification system that detects tree species from bark images. Existing bark identification systems rely heavily on massive computing power access, which may be scarce in many locations. Our approach is deployed as a smartphone application that does not require any connection to a database. Its intended use is in a forest, where internet connection is often unavailable. The tree bark identification is expressed as a bark image classification task, and it is implemented as a convolutional neural network (CNN). This research focuses on developing light-weight CNN models through knowledge distillation. Overall, we achieved 96.12% accuracy for tree species classification tasks for ten common tree species in Indiana, USA. We also captured and prepared thousands of bark images—a dataset that we call Indiana Bark Dataset—and we make it available at https://github.com/wufanyou/DBID.

Automatic species identification has the potential to improve the efficacy and automation of wood processing systems significantly. Recent advances in deep learning allowed for the automation of many previously difficult tasks, and in this paper, we investigate the feasibility of using Deep Convolutional Neural Networks (CNNs) for hardwood lumber identification. In particular, we tested two highly effective CNNs (ResNet-50 and DenseNet-121) as well as lightweight MobileNet-V2. Overall, we achieved 98.2% accuracy for 11 common hardwood species classification tasks.

We investigate how to effectively and efficiently embed users' personalized travel behaviors to vectors in this paper. Based on an example scenario of travel mode choice in intelligent transportation system, three data structures representing users' travel behaviors are defined, namely heterogeneous graph of users' travel behaviors, user travel behavior k-partite graph, and personalized user travel behavior sentence set. This paper systematically analyzes the principle of existing methods and provides intuitions for the problem of learning travel behavior representation in intelligent transportation system. Then we propose the Behavior2vector, which is an improved method tailored for embedding users' personalized travel behaviors to vectors. In our experiments, we design a travel mode choice model based on machine learning, which uses both hand-crafted basic features and embedded vector features. We further quantify the impact of various factors on travel mode choice and use travel big data to test the hypothesis of traffic assignment models, e.g., travelers always choose the path with the shortest path. In addition, we also compared with the existing graph embedding methods and essentially discussed their advantages and disadvantages.

Using historical data to help route planning is significant since the real world is complicated, and the quality of a route is not only defined by its theoretical cost. This report proposed a two-step method that involved learning history and performing classic solutions to vehicle routing problems (VRP). Specifically, the task of the Last Mile Routing Research Challenge is decomposed into two steps: 1) predicting zone level sequence and 2) perform VRP within a single zone. Our method reaches 0.042 locally based on a train test split. The code can be found https://github.com/wufanyou/TLab-Last-Mile.

The IARAI Traffic4cast competition at NeurIPS 2019 showed that neural networks can successfully predict future traffic conditions 15 minutes into the future on simply aggregated GPS probe data in time and space bins, thus interpreting the challenge of forecasting traffic conditions as a movie completion task. U-nets proved to be the winning architecture then, demonstrating an ability to extract relevant features in the complex, real-world, geo-spatial process that is traffic derived from a large data set. The IARAI Traffic4cast challenge at NeurIPS 2020 build on the insights of the previous year and sought to both challenge some assumptions inherent in our 2019 competition design and explore how far this neural network technique can be pushed. We found that the prediction horizon can be extended successfully to 60 minutes into the future, that there is further evidence that traffic depends more on recent dynamics than on the additional static or dynamic location specific data provided and that a reasonable starting point when exploring a general aggregated geo-spatial process in time and space is a U-net architecture.

The problem of the effective prediction for large-scale spatio-temporal traffic data has long haunted researchers in the field of intelligent transportation. Limited by the quantity of data, citywide traffic state prediction was seldom achieved. Hence the complex urban transportation system of an entire city cannot be truly understood. Thanks to the efforts of organizations like IARAI, the massive open data provided by them has made the research possible. In our 2020 Competition solution, we further design multiple variants based on HR-NET and UNet. Through feature engineering, the hand-crafted features are input into the model in a form of channels. It is worth noting that, to learn the inherent attributes of geographical locations, we proposed a novel method called geo-embedding, which contributes to significant improvement in the accuracy of the model. In addition, we explored the influence of the selection of activation functions and optimizers, as well as tricks during model training on the model performance. In terms of prediction accuracy, our solution has won 2nd place in NeurIPS 2020, Traffic4cast Challenge.

Recent advances show that deep neural networks are not robust to deliberately crafted adversarial examples which many are generated by adding human imperceptible perturbation to clear input. Consider l2 norms attacks, Project Gradient Descent (PGD) and the Carlini and Wagner (C&W) attacks are the two main methods, where PGD control max perturbation for adversarial examples while C&W approach treats perturbation as a regularization term optimized it with loss function together. If we carefully set parameters for any individual input, both methods become similar. In general, PGD attacks perform faster but obtains larger perturbation to find adversarial examples than the C&W when fixing the parameters for all inputs. In this report, we propose an efficient modified PGD method for attacking ensemble models by automatically changing ensemble weights and step size per iteration per input. This method generates smaller perturbation adversarial examples than PGD method while remains efficient as compared to C&W method. Our method won the first place in IJCAI19 Targeted Adversarial Attack competition.

We present a probabilistic model for network growth with preferential attachment and self-attractivity. Instead of connecting to a predetermined number of existing nodes, any new node makes a finite number of attempts to connect to previous ones, any trial having a finite probability of success. We find a percolation phase transition which significantly differs from that of related models. Node degree distribution appears scale-free for large degrees, the exponent depending on system connectivity. Cluster size distribution becomes scale-free at the phase transition. Cluster diameter in percolating clusters increases logarithmically with cluster size, but becomes a power-law function at the phase transition. Counted number of boxes becomes exponentially reduced with box size, indicating non-fractal cluster geometry in the connected phase. The geometry however becomes fractal at the phase transition, with power-law exponent 2.