OPTIMIZING SUPPLY CHAIN ROUTES BY MACHINE LEARNING AND OPEN CITYMAP

ABSTRACT

This paper presents a comprehensive study on optimizing supply chain management within contemporary business frameworks, emphasizing the integration of strategic elements such as goods distribution and timely delivery. As modern businesses face increasing demands for efficiency and rapid response due to time-sensitive goods, traditional supply chain models are reevaluated to enhance reliability and reduce dependency on third-party intermediaries. Our research introduces a robust algorithm designed to streamline the process by synchronizing production schedules with distribution strategies and delivery timelines, thereby ensuring that all stages from production to customer delivery are effectively managed. This integration not only addresses the logistics of product flow and vehicle capacity constraints across multi-stage networks but also aligns with cost minimization and efficiency optimization goals. Additionally, the paper compares the efficacy of three machine learning models—Convolutional neural network, K-nearest neighbors, and Random Forest—in optimizing route planning for delivery, highlighting that while Random Forest shows superior performance on map-related inputs, K-nearest neighbors and Random Forest exhibit similar accuracy with time-based data. As a result, this study reinforces the pivotal role of strategic supply chain management in meeting contemporary market demands and enhancing customer satisfaction.

АННОТАЦИЯ

В этом документе представлено комплексное исследование по оптимизации управления цепочками поставок в рамках современного бизнеса с акцентом на интеграцию стратегических элементов, таких как распределение товаров и своевременная доставка. Поскольку современные предприятия сталкиваются с растущими требованиями к эффективности и быстрому реагированию из-за срочных поставок товаров, традиционные модели цепочек поставок пересматриваются для повышения надежности и снижения зависимости от сторонних посредников. В нашем исследовании представлен надежный алгоритм, предназначенный для оптимизации процесса путем синхронизации графиков производства со стратегиями распределения и сроками поставок, тем самым гарантируя эффективное управление всеми этапами от производства до доставки заказчику. Такая интеграция не только решает проблему логистики продуктовых потоков и ограничений пропускной способности транспортных средств в рамках многоступенчатых сетей, но и соответствует целям минимизации затрат и оптимизации эффективности. Кроме того, в документе сравнивается эффективность трех моделей машинного обучения — Конволюционной нейронной сети, K-ближайших соседей и Случайного леса — в оптимизации планирования маршрута доставки, подчеркивая, что, хотя Случайный лес демонстрирует превосходную производительность при вводе данных, связанных с картой, K-ближайших соседей и Случайный лес демонстрируют аналогичную точность при обработке данных, основанных на времени. В результате это исследование подтверждает ключевую роль стратегического управления цепочками поставок в удовлетворении современных требований рынка и повышении удовлетворенности клиентов.

Keywords: Supply Chain Management; Goods Distribution; Timely Delivery; Efficiency Optimization; Machine Learning; Route Planning; Reliability.

Ключевые слова: Управление цепочками поставок; Распределение товаров; Своевременная доставка; Оптимизация эффективности; Машинное обучение; Планирование маршрута; Надежность.

Introduction

Supply chain management has become a pivotal component in the architecture of modern business operations. As products traverse through intricate stages from order taking to warehouse delivery, the efficiency of these processes is paramount. Effective management of these stages involves not only optimizing the flow of goods from suppliers to consumers but also enhancing coordination among manufacturers, distributors, and retailers. This harmonization is crucial for reducing costs, speeding up operations, and ultimately, enhancing overall performance. Key to this management is the optimization of production and inventory processes, aimed at maximizing profitability while minimizing costs and risks.

Further, the evolution of supply chain strategies extends to the timing of raw material supply across multi-stage networks, highlighting the critical nature of product life and delivery schedules. Time-sensitive goods, for instance delivered by Wolt services, require swift supply chain processes to preserve product quality and meet consumer demands promptly. This urgency drives the integration of tasks within the supply chain, potentially eliminating the need for third-party intermediaries. Such integration can lead to cost savings and increased reliability, ensuring that products are delivered efficiently and meet customer satisfaction. This paper contributes to the body of knowledge by presenting a comprehensive model that optimizes these elements, reinforcing the importance of strategic supply chain management in contemporary business environments.

The distribution problem encapsulates crucial logistics activities including production planning, goods distribution, and customer delivery. The complexities and nuances of this problem will be elaborated in subsequent sections.

Literature Review

At this juncture, it is pertinent to explore existing literature to understand the problem's relevance and the progress made in this domain. This issue is significantly impactful on societal operations, thus garnering extensive global research interest [5].

For instance, one study [10] tackles the integration of production, inventory, and distribution through the application of genetic algorithms to develop several heuristic solutions. This research simplifies the logistics model by assuming that all customers reside relatively close to each other compared to their distance from the production site, leading to a fixed shipping cost per customer rather than a comprehensive routing challenge. Another approach [6] simplifies the transportation component by consolidating finished goods into batches. These batches are transported either by the company's vehicles at a fixed cost per customer or by third-party carriers, where costs are both fixed and variable, dependent on the batch size. This setup allows each vehicle to serve only one customer, sidestepping the need for complex routing algorithms.

In contrast, the study in [14] omits the routing challenge by opting for direct shipments where vehicles are fully loaded and each serves a single customer. Meanwhile, the research detailed in [3] introduces an additional layer to the supply chain, considering suppliers of the raw materials used in production. This comprehensive approach tackles both the minimization of total supply costs and the maximization of service levels at the production stage, positing that reducing sales losses directly enhances service quality. Conversely, the paper [8] does not incorporate raw material suppliers but instead expands the distribution framework to include multiple distributors managing several manufacturing sites, thereby increasing the complexity of the logistics and distribution network.

The research in [1] explores a problem akin to the one discussed in this paper but diverges by introducing a probabilistic model to assess potential risks instead of positing definitive solutions. This model enables the supply chain to be rapidly reconfigured following disruptions, through the analysis of various random scenarios that might impact supply chain efficiency. Meanwhile, the study in [4] details a scenario involving parallel production across multiple factories, with subsequent storage in an unlimited-capacity warehouse before distribution to customers using a single delivery vehicle. To address this setup, the authors employ the insert-and-order algorithm and a genetic algorithm for an Iterated Greedy approach.

In a similar way, [15] investigates a scenario where parts are manufactured in a factory, demanded by customers, and then assembled into final products exclusively at customer locations—reflective of practical scenarios such as those encountered in hardware production industries. A significant deviation in this study from the current research is the inability to store inventory at customer sites [7][16][19]. The authors propose a bifurcated approach: initially determining the production schedule, quantity, sequence, and assigning customers to tentative routes and vehicles based on capacity, albeit with only approximate consideration of transport costs and times [9]. The first phase aims to minimize both production and estimated distribution costs. The second phase utilizes the model to precisely schedule and sequence customer visits on each route, already informed by the distribution of vehicles and routes. Solutions are refined through local search techniques within each sub-problem, aiming for optimal routing that satisfies all constraints.

Having reviewed various scholarly works, it is apparent that this problem remains highly relevant due to its practical implications [12][17]. Effective and efficient solutions can significantly reduce operational costs for companies, enhance customer satisfaction, and potentially boost customer loyalty and attract new business, ultimately increasing profitability. This relevance extends across both large and small companies.

The most of problem formulations in the literature shows varied approaches to handling the numerous variables and constraints inherent in these types of problems. While some researchers simplify by omitting less critical aspects [18][20], this paper introduces a novel problem formulation not previously identified in the literature, along with a fresh approach to solving it. The objectives here are two way: to develop a robust algorithm capable of efficiently solving the integrated distribution problem and to evaluate the effectiveness of dividing the problem into manageable parts as suggested in prior studies, assessing the necessity of such an approach. 

Methods and Materials

Review of previous research papers indicates that the framework of the problem varies, with my research drawing primarily on a city transportation system as its base, augmented by additional constraints to reflect real world complexities more accurately. The defined problem is structured around a complete graph , where  is a set of vertices and   consists of edges. Vertex  represents a locations producing items  , and vertices  represent customers. Vehicles are denoted by  , all identical with a capacity . The scenario is assessed over multiple time periods . During each period , each vehicle  can undertake just one route, limited by length . Each customer is served by only one vehicle per period, forbidding the division of orders among multiple vehicles [2].

The producers’ output is confined by its capacity , with production time for item  given by . Production costs for item  are , with setup costs  if item  is produced in period . It's permissible to keep finished goods at the factory for subsequent delivery, incurring a storage cost  per period for item . Realistic constraints on warehouse storage stipulate that item  cannot be stored below  or exceed . Demand for item  from customer  during period  is , allowing the possibility to supply excess goods to a customer for future needs, thereby adjusting later deliveries, with  as the cost for storing item  with customer . Each item  and each customer  have defined minimum  and maximum  storage volumes. Transportation costs include a fixed  if vehicle  is used in period  and a variable  for traveling from node  to node .

Variable Definitions:

 - quantity of product  produced in period ;

 - inventory of product  at customer  at the end of period ;

 - inventory of product  in the warehouse at the end of period ;

1 , if product  is produced in period , 0 otherwise;

 - quantity of product  delivered to customer  via vehicle  in period ;

 - quantity of product  transported on edge  via vehicle  in period ;

1 , if vehicle  passes through edge  in period , 0 otherwise.

We define following objective function:

This objective function minimizes the costs of production, transportation, and storage at the factory and customers, thus leading to an increase in net profit, which is the goal of any company.

Constraints include:

Balancing produced, stored, and delivered goods for the factory:

Balancing produced, stored, and delivered goods for customers:

Production limit constraint to not exceed factory capacity:

Ensuring production only if set up occurs:

where  is some large number.

Reflects the flow of goods in storage at customers:

This constraint ensures the balance of transported and stored goods at customer locations.

Reflects the flow of goods in storage at the factory:

This ensures the proper accounting of goods entering and leaving the factory.

Vehicle capacity constraint:

Limits the total volume of products transported per vehicle to prevent overloading.

Route length limitation for each vehicle:

Ensures that the total travel distance or route complexity per vehicle does not exceed a predefined threshold.

Ensuring no more than one route per vehicle per period:

This guarantees that each vehicle is assigned to no more than one route in each period.

Ensures every vehicle returns to the factory at the end of each period's route:

This ensures that for every vehicle that departs from the factory or a customer, there is a corresponding return, balancing the flow of vehicles.

Ensures that each customer is visited by only one vehicle per period:

This ensures that for every vehicle that departs from the factory or a customer, there is a corresponding return, balancing the flow of vehicles.

Inventory level constraints:

This set of constraints ensures that the quantity of product stored at the customer and in the warehouse remains within the specified upper and lower bounds.

Non-negativity and binary conditions:

This group of constraints verifies the feasibility of the variable values. All production, delivery, and transportation quantities must be non-negative, and certain operational decisions are restricted to binary values.

The challenge of the problem involves determining a production plan, distribution plan, and delivery schedule that together satisfy the demand of all customers, minimize the objective function, and comply with all outlined constraints. This approach ensures a comprehensive management of the supply chain from production to delivery, aligning with the strategic goals of cost minimization and efficiency optimization. Reflecting product flow in storage at customers and the factory, vehicle capacity constraints, and route length limitations for each vehicle. Each of these constraints ensures the feasibility of the variable values within the prescribed limits.

Table 1.

Contribution to the overall effectiveness of the supply chain management

Results

Our research plan also led us to explore the field of route planning and congestion control. While vehicle routing problems emerged as a popular topic, its primary focus on commercial vehicle delivery rendered it less pertinent to our specific product distribution challenges. Nevertheless, it offered valuable insights, especially when considering route planning in dynamic traffic conditions.

Furthermore, we considered studies that employed Radio-Frequency Identification for congestion detection. While these studies shared thematic similarities with our project, our approach diverged significantly. Instead of a passive sensing mechanism, our system fosters active communication among all traffic participants, ensuring a more cohesive and responsive traffic management system [11][21].

It's worth noting that our system's route planning feature might draw parallels with established navigation applications like Google (По требованию Роскомнадзора информируем, что иностранное лицо, владеющее информационными ресурсами Google является нарушителем законодательства Российской Федерации – прим. ред.) Maps and Yandex. However, such comparisons only scratch the surface. Unlike conventional navigation apps that prioritize route guidance, our system emphasizes the symbiotic relationship between the application, vehicles, and users. For instance, our system can autonomously detect a car's low fuel level, eliminating the need for manual input. Additionally, our platform facilitates direct vehicular communication, a feature exemplified in the special service vehicles. In essence, while navigation is a pivotal aspect of our system, it's merely one facet of a broader, more integrated solution.

The foundation of our smart traffic system is the comprehensive dataset derived from the NYC Taxi Trip Data. This dataset, accessible to the public via the NYC Open Data website, is segmented into two distinct files representing yellow taxi and green taxi data. In terms of sheer volume, the dataset is expansive, with a size approaching 9 GB in its csv format. It encompasses over 90 million records, spread across 18 columns, offering a detailed snapshot of taxi operations within the city.

Each column in the dataset provides specific insights. Key data points include location details, trip distance, fare amount, passenger count, tax, and more. For the scope of our project, we utilized all the available rows and approximately half of the columns as features. This selective approach does not diminish the dataset's value; in fact, the unused columns hint at the dataset's versatility and potential for diverse applications.

Given the real-time nature of our system, data retrieval speed was paramount. We employed two methods for this purpose: API-based retrieval and direct download. Handling such a voluminous dataset in real-time posed challenges. However, our system was adept at delivering prompt responses, even with the massive data size. This efficiency was achieved by segmenting the data and processing it in manageable chunks. Additionally, we pre-calculated several crucial metrics, such as average travel times and peak traffic hours, ensuring our system's responsiveness aligned with the real-time needs of moving vehicles.

Of the 18 columns, our primary focus was on three: taxi pick-up location, drop-off location, and the corresponding timestamps. These columns are instrumental in gauging traffic congestion levels. Additionally, the fare amount column was pivotal for the taxi driver recommendation segment of our project. To simulate a real-time environment, we extracted and outputted data based on the recorded pick-up and drop-off times. Figure 1 provides a visual representation of our data preprocessing approach and highlights the columns we retained for our analysis. A notable aspect of the dataset is its zone-based location recording. This necessitated a conversion of the actual map data to a zone-centric format, a process we will delve into in subsequent sections.

Figure 1. New York City yellow taxi trip data

We implemented a periodic data fetching mechanism, ensuring real-time relevance while filtering out superfluous columns. Our initial steps involved processing and analyzing the dataset using Jupyter notebooks. This preliminary analysis facilitated the visualization of traffic patterns and enabled route planning between distinct zones. Subsequently, these functionalities were integrated into a Flask application, as shown in Figure 2, serving as our primary server capable of handling HTTP requests. This system was envisioned to autonomously chart routes for connected vehicles interfacing with our platform. A distinctive feature setting our system apart from prevalent map applications, like Yandex taxi, is its capability to factor in internal car attributes, such as fuel levels and vehicle type. Furthermore, our system was adept at optimizing routes for individual vehicles while simultaneously orchestrating broader traffic conditions.

Our initial server was built using the Streamlit framework. However, its tight frontend-backend coupling posed challenges for our connected cars to interface via HTTP requests. This necessitated a shift to the Flask framework, requiring a server rebuild. With traffic visualization and RESTful API operating on separate devices, synchronization became a challenge. We introduced a 'Refresh' button to address this. To cater to multiple clients simultaneously, we integrated MongoDB database and mandated user log-ins, assigning unique IDs to differentiate sessions. Upon accessing the smart traffic system, users are prompted to log in. They are presented with a navigation interface, allowing them to specify parameters like starting point, destination, time, fuel level, and vehicle type. The system then processes this input, fetching relevant data, either via API or locally. Subsequently, it generates tailored maps, reflecting real-time conditions, and returns them to users. The entire redirection process is facilitated through HTTP requests.

Figure 2. A server application using Flask for item pickups

In its essence, our design encapsulates an overall approach to city transport management, ensuring real-time responsiveness, user customization, and efficient route optimization. The system's adaptability and forward-thinking design make it a promising solution for modern urban landscapes. To establish a baseline for our research, we undertook an analysis of taxi data devoid of real-time features. Our primary metric was the total count of pickups and drop-offs for each zone, which we utilized as a proxy for congestion levels. This static analysis served as a reference point to gauge the efficacy of our real-time system.

By calculating hourly pickup and drop-off counts, we discerned the peak traffic hours. Recognizing these peak times is pivotal as it allows our system to preemptively notify vehicles about impending high-traffic periods. Our findings indicated peak hours stretching from 7 am to 6 pm, with off-peak hours nestled between 3 am to 5 am. Our approach to visualizing congestion levels hinged on a weighted graph. Initially, weights mirrored standard traffic conditions. As real-time data streamed in, we computed the time taken between zone pairs. Any deviation from the "standard time" flagged a zone as congested. This deviation was then scaled to a specific congestion level and updated in the graph, which subsequently informed our route optimization algorithm.

Our system's spatial component was realized through two methodologies: a weighted directed graph and a geojson file, as shown in Figure 3. The absence of adjacency data in online maps necessitated manual labeling. For each adjacent zone pair, we extracted 500,000 records from our dataset, computing average transit times. These averages formed our congestion standard. We encapsulated this data within a class named 'Graph'. Parallelly, for a more visual representation, we sourced a KML file from a government portal, converting it to a GeoJson format using an online tool. Leveraging the Folium library, we crafted an interactive map. Post the extraction of relevant zones, we could manipulate the map, such as color-coding zones. We assigned colors corresponding to varying congestion levels, ranging from free-flowing traffic to highly congested zones.

Figure 3. A geojson file with congested zones

These results underscore the versatility of our smart distribution system. Through a blend of algorithms, real-time data processing, and user-centric features, we've crafted a system that not only optimizes routes but also caters to unique user requirements, marking a significant advancement in the domain of smart traffic systems.

Upon finalizing the weighted directed graph, the search for an optimal route determination algorithm led us to following results as shown in Figure 4. We integrated a mechanism specified for transportation vehicles. This mechanism temporarily adjusted the weights of the map graph, ensuring that regular vehicles were redirected away from the route of any congested zones.

Figure 4. A data file with time delays between zones

The Table 2 compares the effectiveness of three machine learning models—Convolutional neural network, K-nearest neighbors, and Random Forest—judged by their performance on two types of input data for finding fastest routes. It shows that Random Forest outperforms the others on map-related inputs, achieving 82% accuracy, while K-nearest neighbors and Random Forest tie at 87% with time-based inputs. To enhance these models, we consider refining the data preprocessing, and potentially integrating ensemble techniques. Additionally, it’s good to examine whether the complexity of the Convolutional neural network is appropriate for the data, and explore the use of more advanced features or real-time data to capture dynamic patterns in distribution routes. Adjusting these factors may lead to more accurate predictions and a more efficient use of the city’s vehicle data.

Table 2.

Model performance

Our system's current design is based on zone-level data. With more precise data, we could significantly enhance our system's capabilities. However, this would also demand more computational power, leading to a potential trade-off between efficiency and precision. Additionally, while our current system achieves vehicle coordination through a central server, with more hardware support, we could explore direct communication between vehicles or infrastructures, paving the way for a more interconnected and efficient traffic system.

Conclusion

In conclusion, this paper has systematically demonstrated the vital role of advanced supply chain management in modern business landscapes, where the integration of strategic planning and machine learning technologies can significantly elevate operational efficiencies and reliability. By developing and implementing a robust algorithm, this research has addressed complex logistics challenges across the entire spectrum of the supply chain—from production and distribution to the final delivery stages.

The introduction of a sophisticated model that optimizes the synchronization of production schedules, distribution strategies, and delivery timings has proven crucial in meeting the evolving demands of time-sensitive goods and consumer expectations. This optimization ensures not only the adherence to stringent delivery schedules but also the reduction of reliance on third-party intermediaries, thus fostering cost efficiencies and enhancing supply chain resilience. Moreover, the comparative analysis of machine learning models—Convolutional neural network, K-nearest neighbors, and Random Forest—has underscored the potential of these technologies in refining route planning processes. Particularly, the superior performance of the Random Forest model on map-related inputs and its competitive results on time-based inputs highlight the transformative impact of integrating these algorithms within the supply chain frameworks.

By bridging the gap between theoretical approaches and practical implementations, this study contributes significantly to the ongoing discourse in supply chain optimization, offering actionable insights and a scalable model that can be adapted by businesses aiming to improve their logistical operations and overall customer satisfaction. As a result, this research not only reinforces the importance of strategic supply chain management but also sets a benchmark for future explorations in this field.

References:

1. Abualigah L., Hanandeh E.S., Zitar R.A., Thanh C., Khatir S., Gandomi A.H. Revolutionizing sustainable supply chain management: A review of metaheuristics // Engineering Applications of Artificial Intelligence. – 2023.  Vol. 126.
2. Arora S., Brintrup A. How does the position of firms in the supply chain affect their performance? An empirical study // Applied Network Science. – 2021. Vol. 6. №19.
3. Asadujjaman Md., Rahman H.F., Chakrabortty R.K., Ryan M.J. A Memetic Algorithm for Concurrent Project Scheduling, Materials Ordering and Suppliers Selection Problem // Procedia Computer Science. – 2021. Vol. 192. – P.717–726.
4. Badakhshan E., Ball P. A simulation-optimization approach for integrating physical and financial flows in a supply chain under economic uncertainty // Operations Research Perspectives. – 2023. Vol. 10.
5. Baha M.M. Developments of Digital Technologies Related to Supply Chain Management // Procedia Computer Science. – 2023. Vol. 220. P.788–795.
6. Chen Yh. Intelligent algorithms for cold chain logistics distribution optimization based on big data cloud computing analysis // Journal of Cloud Computing – 2020. Vol. 9. №37.
7. El Raoui H., Oudani M., El Hilali Alaoui A. Coupling Soft Computing, Simulation and Optimization in Supply Chain Applications: Review and Taxonomy // IEEE Access. – 2020. Vol. 8. P.31710–31732.
8. Finnah B., Gönsch J., Ziel F. Integrated day-ahead and intraday self-schedule bidding for energy storage systems using approximate dynamic programming // European Journal of Operational Research. – 2022. Vol. 301. №2. – P.726–746.
9. Gopal P.R.C., Rana N.P., Krishna T.V., Ramkumar M. Impact of big data analytics on supply chain performance: an analysis of influencing factors // Annals of  Operations Research. – 2024. Vol. 333. P.769–797.
10. Jiang W. An Intelligent Supply Chain Information Collaboration Model Based on Internet of Things and Big Data // IEEE Access. – 2019. Vol. 7. P. 58324–58335.
11. Jiang W. An Intelligent Supply Chain Information Collaboration Model Based on Internet of Things and Big Data // IEEE Access – 2019. Vol. 7. P.58324–58335.
12. Kehayov M., Holder L., Koch V. Application of artificial intelligence technology in the manufacturing process and purchasing and supply management // Procedia Computer Science. – 2022. Vol. 200. P.1209–1217.
13. Matsui K. Optimal timing of acquisition price announcement for used products in a dual-recycling channel reverse supply chain // European Journal of Operational Research. – 2022. Vol. 300. №2. – P.615–632
14. Nwachukwu C.M., Olofsson E., Lundmark R., Wetterlund E. Evaluating fuel switching options in the Swedish iron and steel industry under increased competition for forest biomass // Applied Energy. – 2022. Vol. 324.
15. Rajesh R. Optimal trade-offs in decision-making for sustainability and resilience in manufacturing supply chains // Journal of Cleaner Production. – 2021. Vol. 313.
16. Salamai A., Hussain O. K., Saberi M., Chang E., Hussain F. K. Highlighting the Importance of Considering the Impacts of Both External and Internal Risk Factors on Operational Parameters to Improve Supply Chain Risk Management // IEEE Access. – 2019. Vol. 7. P.49297–49315.
17. Salçuk, K., Şahin, C. A novel multi-objective optimization model for sustainable supply chain network design problem in closed-loop supply chains // Neural Computing & Applications. – 2022. Vol. 34. P.22157–22175
18. Xia Y., Li C. Robust Control Strategy for an Uncertain Dual-Channel Closed-Loop Supply Chain With Process Innovation for Remanufacturing // IEEE Access. – 2023. Vol. 11. P.97852–97865.
19. Zhao W., Wang D. Simulation-Based Optimization on Control Strategies of Three-Echelon Inventory in Hybrid Supply Chain With Order Uncertainty // IEEE Access. – 2018. Vol. 6. P.54215–54223.
20. Zhiwen Z., Xiaoying Y., Yujun X., Limin G., Junxing L. Supply Chain Logistics Information Quality Evaluation From Just-in-Time Perspective // IEEE Access. – 2020. vol. 8. P.105728–105743.
21. Zhiwen Z., Yujun X., Junxing L., Limin G., Long W. Supply Chain Logistics Information Collaboration Strategy Based on Evolutionary Game Theory // IEEE Access. – 2020. Vol. 8. P.46102-46120.