Tích hợp cơ chế đấu giá và các giải thuật định tuyến vào bài toán giao nhận hàng hóa

ĐẠI HỌC QUỐC GIA TP HỒ CHÍ MINH TRƯỜNG ĐẠI HỌC BÁCH KHOA TRẦN NGỌC BẢO DUY TÍCH HỢP CƠ CHẾ ĐẤU GIÁ VÀ CÁC GIẢI THUẬT ĐỊNH TUYẾN VÀO BÀI TỐN GIAO NHẬN HÀNG HĨA Chun ngành: Khoa học Máy tính Mã số: 8480101 LUẬN VĂN THẠC SĨ TP HỒ CHÍ MINH, tháng năm 2020 LỜI CÁM ƠN Đầu tiên, xin gửi lời cám ơn chân thành đến PGS TS Phạm Trần Vũ cung cấp ý tưởng đề tài, hướng dẫn đọc, nghiên cứu tài liệu tận tình hướng dẫn suốt trình học tập nghiên cứu để hồn thành luận văn tốt nghiệp Tơi xin cám ơn đến thầy cô khoa Khoa học Kỹ thuật Máy tính, mơn Khoa học Máy tính - nơi công tác tạo điều kiện giúp tơi vừa hồn thành cơng việc vừa có thời gian để nghiên cứu để hồn thành nội dung luận văn Để có kết ngày hôm nay, ghi nhớ công ơn gia đình, cha mẹ động viên tơi vượt qua khó khăn, tiếp sức cho tơi vật chất tinh thần Bên cạnh đó, tơi muốn gửi gắm lời cảm kích đến tập thể bạn sinh viên, cán Đoàn Thanh niên - Liên chi Hội Sinh viên Khoa Khoa học Máy tính giúp đỡ tơi thời gian làm luận văn cơng tác điều hành để tơi hồn thành công việc học tập công tác mà Khoa trường giao phó Mặc dù cố gắng hoàn thành luận văn chắn nội dung khơng tránh khỏi thiếu sót Tơi mong nhận thơng cảm, góp ý tận tình bảo quý Thầy cô bạn Tp Hồ Chí Minh, tháng năm 2020 i TĨM TẮT LUẬN VĂN Ngày với phát triển mạnh mẽ ngành Logistics bùng nổ mặt khoa học công nghệ - đặc biệt công nghệ thông tin liên lạc, nhu cầu vận chuyển hàng hóa khu vực kinh doanh vừa nhỏ, kinh doanh dạng hộ gia đình ngày lớn Họ chủ yếu không tự xây dựng hệ sinh thái giao nhận riêng chủ yếu dựa vào đơn vị vận tải thứ ba, mà lựa chọn tập trung chủ yếu vào ứng dụng giao nhận trực tuyến phát triển dịch vụ thời gian gần đầy Tuy nhiên, đơn giá giao nhận phụ thuộc lớn vào bên cung cấp dịch vụ (trung gian người giao hàng người cần giao hàng) Xuất phát từ nhu cầu thực tế đó, đề tài xây dựng theo hướng tích hợp chế đấu giá - chế xác định giá phổ biến từ lâu để điều chỉnh hành vi cung cấp dịch vụ theo hướng có lợi cho người giao hàng người cần giao hàng Đề tài trình bày tập trung thành phần, phần thứ nghiên cứu tốn góc độ tốn học, từ đưa mơ hình để lĩnh vực liên quan khác tham khảo phát triển Trong phần này, tác giả thử nghiệm đánh giá mơ hình đưa chế vét cạn để tìm lời giải tối ưu toán Phần thứ hai phần nghiên cứu giải thuật kinh nghiệm, giải thuật tiến hóa để giải nhanh đưa kết tốt thời gian hợp lý cho toán Báo cáo luận văn tổ chức thành chương Chương trình bày tổng quan đề tài, mục tiêu, phạm vi nghiên cứu ý nghĩa đề tài Chương trình bày kiến thức tảng, sở hình thành giải đề tài Chương trình bày mơ hình tốn học thử nghiệm Chương 4, trình bày cách áp dụng giải thuật tiến hóa để giải tốn Trong chương 6, tác giả tổng kết đưa hướng phát triển cho đề tài ii ABSTRACT Nowadays, with the strong development of Logistics industry and the explosion of technology - especially as the information technology, the demand for goods transport in the small and medium-sized business is increasing significantly They not primarily build their own delivery system and utilize heavily a third-party transport service, but most of the choice is focused primarily on highly-delivered online delivery applications However, the unit price of delivery depends greatly on the service provider (the intermediary between the shipper and the delivery man) Based on these demands, our study was built with integrating the auction mechanism - a pricing mechanism that has been popular for a long time to adjust service delivery behavior with many benefits The topic is presented in two parts, the first part is to study the problem from a mathematical perspective, thereby providing a model for other related fields to refer and develop In this section, I also test the model and offer exhausted mechanism to find the optimal solution of the problem The second part is the study of heuristic and evolutionary algorithms to solve problems quickly and give good results in a reasonable time for the problem This report is divided into chapters Chapter presents an overview of the topic, objectives, scope of the study and the meaning of the topic Chapter presents background knowledge, foundations and problem solving Chapter covers mathematical modeling and experiments Chapters and describe how to apply evolutionary algorithms to solve problems In chapter 6, summarizing and presenting development directions to the topic will be demonstrated iii LỜI CAM ĐOAN Tơi xin cam đoan cơng trình nghiên cứu riêng Số liệu, kết nêu luận văn trung thực có nguồn gốc rõ ràng Các tài liệu tham khảo trích dẫn cụ thể trình bày phần phụ lục Các kết nghiên cứu thực hướng dẫn giảng viên hướng dẫn PGS TS Phạm Trần Vũ HỌC VIÊN THỰC HIỆN iv Mục lục LỜI CÁM ƠN i TÓM TẮT LUẬN VĂN ii LỜI CAM ĐOAN iv Danh sách hình vẽ viii Danh sách bảng x Chương GIỚI THIỆU ĐỀ TÀI 1.1 Bối cảnh đề tài 1.1.1 Sự phát triển mạnh mẽ ngành Logistic 1.1.2 Sự bùng nổ ứng dụng giao nhận tảng di động 1.1.3 Các cơng trình nghiên cứu liên quan 1.2 Mục tiêu, giới hạn đối tượng nghiên cứu đề tài 11 1.2.1 Mục tiêu đề tài 11 1.2.2 Giới hạn phạm vi đối tượng nghiên cứu 12 1.3 Ý nghĩa khoa học thực tiễn đề tài 12 1.3.1 Ý nghĩa khoa học 12 1.3.2 Ý nghĩa thực tiễn 13 1.4 Cấu trúc luận văn 13 Chương CƠ SỞ KHOA HỌC 15 2.1 Cơ sở kinh tế học chế đấu giá 15 2.1.1 Khái niệm đấu giá 15 2.1.2 Các phương thức đấu giá phổ biến 16 2.1.3 Ưu điểm đấu giá trực tuyến 17 2.1.4 Ý nghĩa việc đấu giá chất toán 18 2.2 Các phương pháp tối ưu hóa 19 2.2.1 Nhóm giải thuật tìm kiếm lân cận (local search) 19 2.2.2 Thành tựu tính tốn tiến hóa (Evolutionary Computation) 24 v Chương MƠ HÌNH BÀI TỐN VÀ THỬ NGHIỆM 28 3.1 Mơ hình tốn học toán 28 3.2 Thử nghiệm đánh giá 36 3.2.1 Phương pháp lấy tập liệu đánh giá 36 3.2.2 Các tiêu chí đánh giá 46 3.2.3 Thử nghiệm mơ hình toán học 46 3.3 Lấy lời giải với chế vét cạn toán 52 3.3.1 Backtracking 52 3.3.2 Forward-checking backtracking 58 Chương ỨNG DỤNG GIẢI THUẬT DI TRUYỀN VÀO GIẢI QUYẾT BÀI TOÁN 61 4.1 Hình thành khung cho giải thuật di truyền dựa đặc tính tốn 61 4.1.1 Biểu diễn lời giải 61 4.1.2 Khởi tạo quần thể 63 4.1.3 Đánh giá lợi ích cá thể quần thể 65 4.1.4 Toán tử lựa chọn cá thể sinh sản 69 4.1.5 Toán tử lai ghép 70 4.1.6 Toán tử đột biến 73 4.1.7 Phương pháp lựa chọn hệ 76 4.1.8 Một số tham số quan trọng 77 4.2 Thử nghiệm với khung 80 4.2.1 Thử nghiệm với liệu vét cạn 80 4.2.2 Thử nghiệm với liệu lớn 82 4.3 Đưa kinh nghiệm vào giai đoạn giải thuật 83 4.3.1 Kinh nghiệm khởi tạo quần thể 83 4.3.2 Kinh nghiệm lai ghép 85 4.3.3 Kinh nghiệm đột biến 86 4.3.4 Thử nghiệm sau đưa kinh nghiệm 88 vi Chương CẢI TIẾN GIẢI THUẬT DI TRUYỀN 90 5.1 Ngữ cảnh sinh học 90 5.1.1 Nhiễm sắc thể (Chromosome) 90 5.1.2 Thông tin di truyền 92 5.1.3 Quá trình sinh sản 92 5.1.4 Nhận xét chung 93 5.2 Hình thành khung cải tiến cho giải thuật di truyền 95 5.2.1 Biểu diễn lại cá thể 95 5.2.2 Đánh giá lại độ lợi cá thể 96 5.2.3 Quá trình sinh sản 96 5.2.4 Khung chung cho giải thuật di truyền cải tiến 96 5.3 Thử nghiệm với khung cải tiến 97 5.3.1 Thử nghiệm với liệu vét cạn 98 5.3.2 Thử nghiệm với liệu lớn 99 5.4 Đánh giá mức độ cải tiến 99 Chương TỔNG KẾT VÀ HƯỚNG PHÁT TRIỂN 102 6.1 Tổng kết 102 6.1.1 Đánh giá chung 102 6.1.2 Ưu điểm 103 6.1.3 Các mặt phải cải tiến 103 6.2 Hướng phát triển 104 CƠNG TRÌNH NGHIÊN CỨU ĐÃ CÔNG BỐ 105 TÀI LIỆU THAM KHẢO 112 Tài liệu tham khảo 112 vii Chương TỔNG KẾT VÀ HƯỚNG PHÁT TRIỂN 6.1 Tổng kết 6.1.1 Đánh giá chung Nhìn chung, đề tài làm hồn thành mục tiêu đề ra, có hai việc, mơ hình hóa tốn đưa giải thuật di truyền cải tiến để giải tốt tốn Những việc làm tóm tắt công việc cụ thể sau đây: Tìm hiểu tích hợp việc đấu giá cho tài xế vào ứng dụng giao nhận để tránh việc tăng giá đột ngột đơn vị vận tải Xây dựng cung cấp mơ hình thực tiễn cho dịch vụ giao nhận phát triển thời gian gần Xây dựng mô hình tốn học riêng cho tốn giải quyết, góp tốn vào lĩnh vực định tuyến cho xe (vehicle routing problem) Tìm hiểu giải thuật tối ưu thử nghiệm nhiều phương pháp để giải mơ hình Xây dựng liệu ngẫu nhiên sở liệu địa lý thực Phát giải thuật di truyền có khả giải tốn cách dễ dàng 102 Nghiên cứu, khảo sát chứng minh tính đắn số kinh nghiệm giải nhanh giai đoạn GA Nghiên cứu đặc tính sinh học cải tiến giải thuật di truyền 6.1.2 Ưu điểm Luận văn trình bày cơng việc giải đề tài góc nhìn nhiều lĩnh vực vận trù học, giao thơng vận tải khoa học máy tính Cách tiếp cận vấn đề luận văn tổng qt, ứng dụng cho nhiều tốn vận tải tương tự Luận văn đưa bước áp dụng giải thuật di truyền cụ thể tốn, từ dễ dàng cải tiến thêm để đạt kết tốt Các kinh nghiệm để cải thiện giải thuật cụ thể hóa tốn nghiên cứu thực nghiệm đánh giá, phù hợp với thực tiễn Ngoài ra, luận văn đưa giải thuật di truyền cải tiến dựa kiến thức sinh học Giải thuật di truyền không áp dụng tốn mà áp dụng để giải tốn địi hỏi cải thiện tốc độ hội tụ lời giải 6.1.3 Các mặt phải cải tiến Do giới hạn nghiên cứu tập trung đơn hàng lên kế hoạch sẵn nên đáp ứng phần lớn đơn hàng (theo khảo sát) Các đơn hàng nhỏ lẻ (theo thời gian thực) chưa đề tài đề cập cách giải Giải thuật di truyền cải tiến đề tài đưa nêu cách áp dụng toán xét, chưa cụ thể cách áp dụng tổng quát nhiều toán khác 103 6.2 Hướng phát triển Sau nghiên cứu đề tài, tác giả đề xuất số hướng phát triển, cụ thể mặt thực khoa học Về mặt thực tiễn, nội dung đề tài thương mại hóa thành sàn gọi xe tích hợp giao nhận ứng dụng chế đấu giá thành làm chế chủ đạo việc định giá đơn hàng giao nhận Nghiên cứu phát triển thành toán với đơn hàng đến theo thời gian thực Các đơn vị vận tải tiếp tục nghiên cứu áp dụng cho thị trường gọi xe, vốn có quy mơ lớn địi hỏi yêu cầu ngặt nghèo Về mặt khoa học, đề tài thêm cập nhật số ràng buộc khác để trở thành nhánh toán vận chuyển Giải thuật di truyền đề xuất với lời giải lưỡng bội có khả ứng dụng cho nhiều tốn khác nhau, địi hỏi lời giải đưa với tốc độ hội tụ nhanh 104 CƠNG TRÌNH NGHIÊN CỨU ĐÃ CƠNG BỐ [1] D T Ngoc Bao and T V Pham, "Collaborative Vehicle Routing for Auction-based Pickup and Delivery Dispatch Services," 2019 11th International Conference on Knowledge and Systems Engineering (KSE), Da Nang, Vietnam, 2019, pp 1-6, doi: 10.1109/KSE.2019.8919466 105 Collaborative Vehicle Routing for Auction-based Pickup and Delivery Dispatch Services Duy Tran Ngoc Bao Tran Vu Pham Faculty of Computer Science and Engineering Ho Chi Minh City University of Technology, VNU-HCM Ho Chi Minh City, Vietnam duytnb@hcmut.edu.vn Faculty of Computer Science and Engineering Ho Chi Minh City University of Technology, VNU-HCM Ho Chi Minh City, Vietnam ptvu@hcmut.edu.vn Abstract—The surge pricing in ride-hailing applications made customers and partners difficult to decide their bookings Auction-based approaches are often used to optimize the costs of parties involved such as customers, drivers, and service providers and the overall performance of the whole system Balancing the benefits of parties involved is a challenging issue By analyzing current auction-based and fixed-price models for pickup and delivery dispatch systems, this paper identifies sustainable constraints and objectives for adapting the auction-based solutions to today highly competitive market, and formalizes the problem in a three-index mathematical model A modified genetic algorithm with more actual biological features and the ability to abbreviate the converge time is introduced to tackle this problem Computational results on models with optimization solver, original and modified genetic algorithms are included to illustrate the advantages of the proposed solution Index Terms—auction, pickup and delivery problem, vehicle routing, genetic algorithm I I NTRODUCTION In recent years, ride-hailing applications is well-known as the best solution for people to get a ride or delivery in many cities in the world, especially as the Southeast Asian market, in which motorbikes are absolutely popular Due to rapid development, those enterprises become more highly arbitrary in pricing They usually run marketing campaign, give promotion in price to get higher market shares so that the price in some period is not fixed as listed depending on their finance sources However, since bad weather or lack of drivers, the price will be significantly increasing Besides, in some markets, there is only an company which is monopoly in this field and it can make the surge pricing for its customers and partners Consequently, there are many protests against these systems’ policies [12] Obviously, these policies in pricing for participants were not actually free by supply and demand on markets They depends on many sources and conditions in particular markets as we mentioned [1] Therefore, the customers is not actually delighted to these systems because they are easy to know that the temporary in pricing In this paper, our goal is to introduce the solution to eliminate these disadvantages The auction mechanism should 978-1-7281-3003-3/19/$31.00 © 2019 IEEE be selected from a few new approaches to tackle because of the natural of it in markets In general, auction is a successful concept to know the actual price of merchandises when sellers not know and certainly they are decided by the demand of the buyers Moreover, we also introduce a new framework for genetic algorithm which is more actual biological characteristic in the natural environments, which could be applied to deal with this problem faster in convergence II R ELATED WORK There are a few applications integrating auction-based principles in their policies to get the highlight from the world market One of notable representatives is Jugnoo from India While Jugnoo in origin is famous with traditional ride-hailing applications, it goes to Singapore - a country with many startup in this field, by a price reverse bidding system [13] In China market, Didi Dache - a part of Didi Chuxing, has applied an auction-based model to attract the customer and the partner because of free pricing, depending on supply and demand It accommodated the user experience in each driver, then the bid price trusts in it However, China government forced them to stop this model because current policy and law not meet the requirements of these systems [14] Besides, BidPrice in US, AUTO in Singapore and Anterin in Indonesia (powered by HERE map), had utilized auction mechanisms to build a price reverse bidding application as the new approach in this highly arbitrary market Despite the fact that they also have got the great compliment at first, they cannot powerfully develop and emulate to innovate in this field compared to other ongoing applications There are many reasons for this, but balancing the profit of each driver is a huge problem and it is the key to get in touch with partners permanently These auction-based applications shows all bid details of participating drivers consists of price, waiting time, distance from them, so users can select anyone they want and usually they will get the lowest bid price Therefore, if drivers have a marvelous strategy to bid, they should get less requests than others and there are many potential conflicts in these systems [14] Generally speaking, the dispatcher is necessary for them to reduce these advantages and stabilize the environment for partners to get more uniform profit matching with their contributed times In [1], the author points out that the growth of traditional systems had not strongly fluctuating in recent times It needs a new highlight model to treat market into another way This research has proposed an auction-based model to adapt the requirements of passengers and drivers It also has the result on testing on various model and compare to current model to illustrate the performance and benefits for two side of this dispatch model Auction mechanisms was used in another fields a long time ago Many researchers suggests a procedure for this way to apply in the vehicle routing [4] with 5-phase and indicate the way to dispatch effectively The profit sharing method between carriers were mentioned and valuable for us to apply in our objective in model B Coltin and M Veloso in [2] have deployed a fleet of robots to conduct a model of pickup and delivery with auction between them to navigate robots to delivery items in a building with requirement satisfaction and optimal costs In this paper, we try to solve the static version of pickup and delivery problem, that is a part of ride-hailing applications There are all requests, drivers and the bid of each driver on each request before the dispatch system for routing runs Two reasons are given to point out that we should firstly deal with this static version Most users using this service come from small and medium enterprises and they almost should have a plan to deliver merchandise to their customers in the day before The other reason that this version is the base of the dynamic one, because one way to solve effectively is resolve the static version of the problem III P ROBLEM DESCRIPTION An auction-based pickup and delivery problem with timewindows (A-PDPTW) is a derivative of profit-sharing multipickup and delivery problem with time-windows (MPDPTW) [3] Within it, there are V vehicles (a vehicle is used by a delivery man) and R requests (customers) Each delivery man is a counterpart of this serivce as definition, so they not start and finalize their routes at the same depot (like in delivery companies) They have the private starting and ending locations and the time-windows for joining this service On the other hand, the requests have same features in the original MPDPTW with pickup and delivery locations, time-windows to take the entire requirements Consequently, there are some following sets: • S = {2k − 1|1 ≤ k ≤ V}: the set of starting nodes of all vehicles that vehicle k-th has starting node 2k − • E = {2k|1 ≤ k ≤ V}: the set of ending nodes of all vehicles that vehicle k-th has ending node 2k • NV = S ∪ E : the set of all vehicles’ starting and ending nodes • P = {2k − 1|V + ≤ k ≤ V + R}: the set of pickup nodes of all requests that request k-th has pickup node 2(k + V) − • D = {2k|V + ≤ k ≤ V + R}: the set of delivery nodes of all requests that request k-th has pickup node 2(k + V) NR = P ∪D : the set of all requests’ pickup and delivery nodes • V = {k|1 ≤ k ≤ V}: the set of available vehicles The graph G = (N, A) includes the nodes N = NV ∪ NR Each node i ∈ N has the service point of time ti and a timewindow [Ai , Bi ], such that Ai ≤ ti ≤ Bi However, in the reality, only nodes in NR need an adequate time-windows All nodes in S need only the starting time and nodes in E need only the ending time Given time-window, a vehicle can arive at node i, and wait until the opening of time-window to start the pickup or delivery service Morever, a vehicle, which is assigned to a request i, must arrive at pickup node before corresponding delivery location Each arc (i, j) in A = N ×N has two weighted parameters: Tij , Cij represent travel time, driver’s cost travel, respectively Actually, we can omit all of arcs (i, j) in S × E because the optimal routing can not include the direct travel from starting to ending nodes of each request, and obviously, of two different requests With each request, all delivery men have to take a bid with a price Bik , known as the bid price of driver k on request i-th Since the feasible solution assigned a request to only a k vehicle, we can convert this parameter format to Bij for easier k calculation Bij must be the bid price of driver k on request having the starting node i, for all j ∈ N and the rest of them can be set as zero A solution to the A-PDPTW is an ordered feasible assignment of each request to each vehicles of the fleet, which the node’s service point of time is in satisfaction of time constraints The problem can be mathematically modeling with the following variables: 1, if driver k travels arc (i, j) k • xij = 0, otherwise • ti is the service point of time of node i Therefore, the A-PDPTW proposed mathematical model has the following constraints, subject to: • ∀k ∈ V : ∀i ∈ S : ∀i ∈ E : xk2k−1,j = j∈N xkj,2k = (1) xkj,i + = (2) xki,j + = (3) j∈N xki,j = k∈V j∈N k∈V j∈N xkj,i = k∈V j∈N ∀i ∈ NR : k∈V j∈N xki,j = k∈V j∈N ∀i ∈ S, k ∈ V : ∀i ∈ P, k ∈ V : (4) xkj,i+1 (5) xkj,i+1 (6) xkj,i (7) k∈V j∈N xki,j = j∈N j∈N xki,j = j∈N ∀i ∈ NR , k ∈ V : xkj,i = j∈N xki,j = j∈N j∈N ∀i ∈ N, k ∈ V : xki,i = (8) • The income of vehicle k: Ik = i∈N j∈N ∀i ∈ N, j ∈ N : ti + Tij ≤ tj + 1− xki,j M (9) (10) (11) ∀i ∈ NR : ≤ ti ≤ bi (12) ∀i ∈ S ∪ P : ti ≤ ti+1 (13) Constraint (1) guarantees that the driver’s route begins and ends at correct nodes Constraint (2) and (3) make no any travels to starting nodes and from ending nodes, such that there are always one and one travel from starting and to ending ones, whereas constraint (4) makes sure that if any driver enter to any node in pickup and delivery nodes set, they should go out and travel to another Constraint (5), (6) and (7) force that a starting and corresponding ending node or a pickup and corresponding delivery node must be visited by only a vehicle Constraint (8) verifies that there is no loop in a node in this routing From constraint (9), they are all time constraints representing the time-windows Constraint (9) guarantees the travel time between two nodes i and j in routes and simultaneously, eliminates the sub-tour or cycles in solutions M is considered as a big number and equals to max {bi + Tij − aj , 0} for each constraint (9) Constraint (10), (11) and (12) specify time-windows of starting, ending, pickup and delivery nodes, respectively The final constraint determines that the delivery and ending nodes must be serviced after correlative pickup and starting nodes One of this model’s key value is no any permanent objective function to minimize (or maximize) We put in great effort to design the homogeneous decision variables to accommodate more objective The reason we prefer this model is the fact that each enterprise (service) has its confidential strategies to dominate large market We can suggest some parts of objective function: • The total expense of all customers: i∈N jN kV ã (14) xkij ì Cij (15) The total cost of all vehicles: C= i∈N j∈N k∈V • k Bij × The cost of vehicle k: Ck = i∈N j∈N xkij × Cij (18) Fk = Ik − Ck ∀i ∈ E : ti ≤ bj E= (17) The profit of vehicle k: k∈V ∀i ∈ S : ti xkij ã k xkij ì Bij (16) ã The standard deviation of profits: V−1 S= Fi − F¯ i∈V (19) with F¯ = Fi is the mean of all the vehicle’s V i∈V profits Obviously, there are the huge number of goals needs to be optimized depending on private strategies of each enterprise which dispatches the service They can combine these of above objective functions to satisfy their requirements by weighted parameters, such as: O = w1 × E + w2 × C + w3 × i∈V (20) Fi + w4 × S By the point of view and priority of each goal, they can choose the combination of weighted parameters wi to reach their plan and get largest share on highly competitive market Dynamically, they can adjust them over the time to find out the final efficiency 8 (a) By total vehicles profit (b) By total vehicles profit and profit standard deviation routing Fig 1: An example for a optimal routing by different objective functions For example in Figure 1, assume that time-windows are satisfied, if an enterprise sets its goals as maximizing only the total vehicles profit, the router will assign all the requests to only driver It is not balanced in this system for the reason that driver has to travel from source to destination with no load If this problem occurs in a large set of vehicles, many drivers can not cooperate with this enterprise in the long time and choose the others instead So that, with balancing the total vehicle profit and minimizing the profit standard deviation, it helps the routing model to distribute each request to suitable driver, such as in Figure 1b IV T HE PROPOSED ALGORITHM : M -GA A Biological context In biological microcosm, a chromosome is a deoxyribonucleic acid (DNA) molecule, which is a molecule composed of two chains that coil around each other [15] For instance, gene A and a represents the high of a tree Gene A appoints that tree is tall, the rest appoints that tree is short and gene A is stronger than gene a So that, individual with Aa is observed as a tall tree, but aa is a short tree In this example, an individual Aa is specially attractive, because it has gene A and a, but the observable traits is only the tall, saying that it also has the opportunity to inherit another trait to the next generation Generally speaking, each individual in the reality has the hidden charming that cannot be observable B Description In the original genetic algorithm, an individual is considered as a chromosome and also, a feasible solution Taking advantage of this feature biological features we mentioned before, an individual in this paper should be redefined as a pair of feasible solutions, or two chromosomes The fitness of an individual can be set by the higher one of child solutions Therefore, with a single body, there are concurrently the observable and hidden charming fitness It makes the population not only more diversity but also less number of individuals considered The flow of our modified genetic algorithms simulates the process of reproduction in microcosm, following as: Population initialization Fitness calculation Mate selection Chromosome division Chromosome mutation Reproduction Recombination Next generation selection Fig 2: The process of our modified genetic algorithm 1) Population initialization: This is first stage in the whole process to complete this algorithm Like in the original, the population size is carefully studied and selected from the number of drivers and requests It will be a constant parameter and our solution needs to remain it all over the generations By many experiments from small to large instances, we made a mathematical induction to formulate the population size to be efficient in our algorithm and display it in section IV Despite the significance of the population size, describing a solution in an effective way in computing is a hard problem to manage In the whole below experiments, we always build a matrix, which has 2(V + R) columns and V rows The value of each cell is the service point of time at which driver has to work, while if driver is not assigned to this location, it will be a positive infinity ∞ Similarly, we can randomly assign the value of corresponding matrix, such as: 101 ∞ 506 ∞ ∞ 232 ∞ 782 ∞ 296 ∞ 525 377 ∞ 415 ∞ Fig 3: Corresponding solution (chromosome) for Figure Out of time-windows constraints, in this way to describe, we must note that the data in a cell and the next one is symmetric That means if a value is not infinity, the next one should not be an infinity sign and otherwise 2) Fitness calculation: Calculating the fitness is not advanced in our design Each individual in population consists of chromosome, so that the fitness function works two times than same size population in original solution It takes the disadvantages of the whole solution in computing, but we can eliminate it by caching the calculated fitness of all solutions 3) Mate selection: In the original, strategy to select individuals to mating pool has many versions, especially as probability-based random Continuing simulation of mating selection in natural environment, we propose a solution for this problem The foundation of it is the attraction of definitely different individuals and the contingency of ”inbreeding” We divide the population into ”male” and ”female” sets by the mean mean(O) of the fitness Oi Two individuals selected into mating pool are opposite in ”gender” The percentage of attraction between two individuals i and j: Aij = Oj − mean(O) Oi − Oj (21) All the couples into mating pool will be selected based on this percentage 4) Reproduction: This stage includes sub-process: • Division: like the meiosis in real chromosome, before participating in reproduction, each individual need to divide to two of chromosomes • Mutation: in the real world, most of mutation in species usually happens when after the chromosome division in meiosis Receiving the mutation rate as a parameter to select the mutating chromosome, all the changes will be compulsory for keeping compliance in time-windows of each location So that, we divide the mutation in this problem into categories: 1) Changing the current value at available cells 2) Remove a pair of request node to another driver • Recombination: After the mutation, the single chromosomes have to be recombined and re-established a new living 5) Next generation selection: Before generation goes into work, we need to select some best individuals with higher fitness for maintaining by elite size With children after reproducing, we select by the priority whose parents have the advantage fitness for the population V C OMPUTATIONAL EXPERIMENTS In this section, we specify our computational experiments to demonstrate the advantages of our proposed algorithm (called as modified genetic algorithm) compared to the original A Generation of test instances Using the characteristic of instances in [3], we construct a new approach to get instances for validation because that all instances has the same starting and ending depot for all drivers The process can be referred as following: 1) Build the program to check the feasibility of constraints for a dynamic instance of only one driver with the arbitrary number of requests by brute-force strategy Because of one driver, the number of combinations which generated from it is not too large, so that, we will not wait too long for this 2) If any requirement needs V vehicles and R requests, R we will generate the instance of one drivers and V requests, check the feasibility and concatenate them into a complete instance 3) Randomly generate the bid price of each driver to requests The key value of step 2, generate requests for each driver, we have to follow these parameter type and domain in the table: TABLE I: Parameter type and domain for randomizing one driver and requests Name Nodes Starting time s Ending time e Pickup time pi Delivery time di Type Coordinate (x, y) Integer Integer Integer Integer Domain x, y ∈ [−50, 50] [0, 1440] [0, 1440] [s, e] [pi , e] Before bid price generation, we had survey delivery service’s price which are well-known in the market Then, we selected the formulation to calculate a request in current price: mp = 0.64 + (d − 2) × 0.21 (22) with 0.64 in US dollar is the opening price for the first kilometers and 0.21 is the price per kilometer after them while d is the distance from the pickup to delivery location We also use the Euclidean for easy distance calculation for because of focusing on the distance in place of the location The expense for driver traveling is $US 0.03 per length unit Therefore, the bid price for a driver and a request are randomly generated from the price-window cost and market price In the test results below, we use the symbol to express an instance: [X] [V ] [R] with: • X is the categorized instance, having one of three values: – S: a location to start and another to end for all drivers – L: a location to start and end for each driver – D: different locations to start and end for all drivers • V: the number of drivers (vehicles) • R: the number of requests B Test results All experiments were conducted on HP Z6 G4 Workstation equipped with Intel 4108 Xeon1.8 8C and 32 Gigabyte RAM The operating system of this workstation is Ubuntu 18.04 LTS The AMPL (A Mathemetical Programming Language) with open-source GLPK 4.32 were used to solve the above MILP model The original and modified genetic algorithm was implemented in Python 3.7.3 using the development IDE Pycharm Edu 2018.3 GLPK only solves linear model, so the objective function used in experiments is to minimize: O = 1.3 × E + 1.1 × C (23) It means that the goal of enterprise, for example, using this objective is minimizing the total expense of all customers and total cost of all vehicles, but the priority of the first is higher It helps them to reduce the expense for customer and to balance with vehicles, avoid the constant assignment of lowest bid price In table II, we displayed the result when using the GLPK solver with model in Section II to find out the optimal solution There is no any time limit in these experiments The solution solved by GLPK is the best, but the run time is too large to handle the requirements in actual applications Comparing results by the original and our modified genetic algorithm, we realized that the solutions by m-GA were slightly decreasing and nearly closer to the best solution than the o-GA ones and the run time was strongly reducing and suitable for certain cases These gap percentages are acceptable due to small difference in values compared with best solution Generally speaking, two evolutionary approaches could be utilized in this problem to get common results However, these approaches are random-based so we carried on the continuous experiments to get the average outcomes for comparison The results was shown in the Figure by instance D 100 500 Although the final results is nearly the same, the m-GA needed only approximately 50 generations to reach the o-GA solution, compared to nearly 200 generations The converge rate of m-GA were illustrated through generates TABLE II: Computational Result with original (o-GA) and modified (m-GA) genetic algorithm Instance S 10 L 10 D 10 S 10 50 L 10 50 D 10 50 S 30 150 L 30 150 D 30 150 S 50 250 L 50 250 D 50 250 S 100 500 L 100 500 D 100 500 GLPK Solver Best solution Time (seconds) 49.9 48.2 68.4 103.9 57.8 152.4 248.1 1882.5 466.5 1826.5 541.4 2452.0 1464.4 2444.1 1485.8 2886.3 1614.7 2893.3 2262.3 7552.4 2434.6 7528.1 2387.2 7612.0 4634.2 14280.1 4711.1 15442.6 4845.1 16223.7 Original genetic algorithm Solution Gap (%) Time (seconds) 52.6 5.4 2.8 73.2 7.0 5.1 63.0 9.1 8.9 267.5 8.7 13.3 507.6 8.8 17.1 582.6 7.6 16.7 1632.4 11.5 36.6 1572.1 5.8 35.1 1731.1 7.2 31.8 2563.9 13.3 50.1 2867.1 17.7 66.3 2779.1 16.4 89.5 5264.9 13.6 154.6 5173.6 9.8 182.0 5353,9 10.5 198.3 R EFERENCES ·104 m-GA o-GA Objective value 1.2 0.8 0.6 Modified genetic algorithm Solution Gap (%) Time (seconds) 50.1 0.4 2.6 73.1 6.8 5.5 62.9 8.8 8.4 262.9 6.0 14.9 482.6 3.5 18.3 564.7 4.3 18.1 1591.7 8.7 38.4 1544.3 3.9 34.7 1671.8 3.5 34.7 2510.4 11.0 51.7 2760.7 13.4 62.1 2768.4 15.9 80.6 5014.0 8.2 131.3 4763.5 1.1 145.1 5013.0 3.5 134.1 100 200 300 400 500 Generations Fig 4: Objective values through generations in instance D 100 500 in the rest of instances Based on that, we certainly used this new approach in GA to get the improvement of run time VI C ONCLUSIONS This paper presents a mathematical model for auction-based multi-pickup and delivery problems, which are applicable to popular ride-hailing applications as a new principle to attract the customers The key value of our model is to adapt to many goals which depends on each enterprise’s marketing strategies Thus, if they design great goals, it will bring customers and partners to them We also proposed a modified genetic algorithms with view of the microcosm and get a comparison to the original one The performance in computing is slightly improving but opens the new approach to solve optimization problem In future work, we will develop the model for dynamic version of this problem and applied to an applications in reality We are also going to build a variant in riding taxi system [1] P.D Kusuma, Auction Based Dispatch Model in Online Motorcycle Taxi System, Journal of Theoretical and Applied Information Technology, vol 96(18), 2018, pp 6134-6148 [2] B Coltin and M Veloso, ”Online pickup and delivery planning with transfers for mobile robots,” 2014 IEEE International Conference on Robotics and Automation (ICRA), Hong Kong, 2014, pp 5786-5791 doi: 10.1109/ICRA.2014.6907709 [3] Naccache, S., Cote, J F., Coelho, L C., 2018 The multi-pickup and delivery problem with time windows European Journal of Operational Research 269 [4] Gansterer, M., Hartl, R F., 2016a Combinatorial auctions in collaborative vehicle routing IFORS News (10(4)), 1516 [5] Dai, B., Chen, H., Yang, G., 10 2014 Price-setting based combinatorial auction approach for carrier collaboration with pickup and delivery requests Operational Research 14 (3), 361386 [6] M G S Furtado, P Munari, and R Morabito, Pickup and delivery problem with time windows: A new compact two-index formulation, Operations Research Letters, vol 45, no 4, pp 334 341, 2017 [7] Jung, Soojung & Haghani, Ali (2000) Genetic Algorithm for a Pickup and Delivery Problem with Time Windows Transportation Research Record 1733 1-7 10.3141/1733-01 [8] Li B., Lau H.C (2017) A Combinatorial Auction for Transportation Matching Service: Formulation and Adaptive Large Neighborhood Search Heuristic In: Bekta T., Coniglio S., Martinez-Sykora A., Vo S (eds) Computational Logistics ICCL 2017 Lecture Notes in Computer Science, vol 10572 Springer, Cham [9] B Coltin and M Veloso, ”Online pickup and delivery planning with transfers for mobile robots,” 2014 IEEE International Conference on Robotics and Automation (ICRA), Hong Kong, 2014, pp 5786-5791 [10] Dumas,Y., J Desrosiers and F Soumis 1991 The Pick-up and Delivery Problem With Time Windows Eur J Opnl Res 54, 7-22 [11] Lu, Q., & Dessouky, M (2004) An Exact Algorithm for the Multiple Vehicle Pickup and Delivery Problem Transportation Science, 38(4), 503-514 [12] Todayonline.com, Numerous complaints about Grab over increased fares and commission fees, says competition watchdog, 2018 Available: https://www.todayonline.com/singapore/numerous-complaints-aboutgrab-over-increased-fares-and-commission-fees-says-competition [Accessed: June 08, 2019] [13] Channelnewsasia.com, Indian ride-hailing firm Jugnoo to enter Singapore with price-bidding system Available: https://www.channelnewsasia.com/news/singapore/indian-ride-hailingfirm-jugnoo-to-enter-singapore-with-price-10168246 [Accessed: June 08, 2019] [14] sina.com.cn, ŁŁ Available: http://news.sina.com.cn/o/2017-01-22/docifxzutkf2327804.shtml [Accessed: June 08, 2019] [15] Dela Pea Jr, Renato & Gracilla, Daniel & Pangilinan, Christian (2016) GENERAL BIOLOGY Tài liệu tham khảo [1] Kusuma, P.D (2018) Auction based dispatch model in online motorcycle taxi system Journal of Theoretical and Applied Information Technology 96 6134-6148 [2] Coltin, Brian & Veloso, M (2013) Online pickup and delivery planning with transfers for mobile robots AAAI Workshop - Technical Report 8-13 [3] Salma Naccache, Jean-Franc¸ois Cơté, Leandro C Coelho, The multipickup and delivery problem with time windows, European Journal of Operational Research, Volume 269, Issue 1, 2018, Pages 353-362, ISSN 0377-2217, https://doi.org/10.1016/j.ejor.2018.01.035 [4] Dai, Bo & Chen, Haoxun (2011) Price-setting based Combinatorial Auction Approach for Carriers’ Collaboration in Less than Truckload Transportation ICAART 2011 - Proceedings of the 3rd International Conference on Agents and Artificial Intelligence 407-413 [5] Gansterer, M., Hartl, R F., 2016a Combinatorial auctions in collaborative vehicle routing IFORS News (10(4)), 15–16 [6] Furtado, Maria & Munari, Pedro & Morabito, Reinaldo (2017) Pickup and delivery problem with time windows: A new compact two-index formulation Operations Research Letters 45 10.1016/j.orl.2017.04.013 [7] Mes, Martijn & van der Heijden, Matthieu Schuur, Peter (2010) Look-ahead strategies for dynamic pickup and delivery problems OR Spectrum 32 395-421 10.1007/s00291-008-0146-3 112 [8] Cai, Yan & Tang, Ya & Yang, Qi (2015) An Improved Genetic Algorithm for Multi-Depot Heterogeneous Vehicle Routing Problem with Simultaneous Pickup and Delivery Time Windows Applied Mechanics and Materials 738-739 361-365 10.4028/www.scientific.net/AMM.738-739.361 [9] Li, Baoxiang & Lau, Hoong (2017) A Combinatorial Auction for Transportation Matching Service: Formulation and Adaptive Large Neighborhood Search Heuristic 127-142 10.1007/978-3-319-684963_9 [10] Dumas, Yvan & Desrosiers, Jacques & Soumis, Franc¸ois (1991) The pickup and delivery problem with time windows European Journal of Operational Research 54 7-22 10.1016/0377-2217(91)90319-Q [11] Lu, Q., & Dessouky, M (2004) An Exact Algorithm for the Multiple Vehicle Pickup and Delivery Problem Transportation Science, 38, 503-514 [12] Cormen, T., Leiserson, C.E., Rivest, R.L., Introduction to Algorithms, The MIT Press, 1997 [13] K Rosen Discrete Mathematics and its Applications McGraw Hill, 1999 [14] Goldberg, D.E., Holland, J.H Genetic Algorithms and Machine Learning Machine Learning 3, 95–99 (1988) [15] Google, e-Conomy SEA 2019 - Swipe up and to the right: Southeast Asia’s $100 billion Internet economy, 2019 Truy cập lần cuối: 08/6/2019 https://tinyurl.com/sal3vbm 113 [16] Mordor Intelligence, Vietnam courier, express, and parcel (CEP) market - growth, trends, and forecast (2020 - 2025), 2019 Truy cập lần cuối: 08/6/2019 https://www.mordorintelligence.com/industryreports/vietnam-courier-express-and-parcel-cep-market [17] Mordor Intelligence, Courier, express, and parcel (CEP) market - growth, trends, and forecast (2020 - 2025), 2019 Truy cập lần cuối: 08/6/2019 https://www.mordorintelligence.com/industryreports/courier-express-and-parcel-cep-market [18] Marketsandmarkets.com, Ride Sharing Market by Type (Ehailing, Station-Based, Car Sharing Rental), Car Sharing (P2P, Corporate), Service (Navigation, Payment, Information), Micro-Mobility (Bicycle, Scooter), Vehicle Type, and Region - Global Forecast to 2025 Truy cập lần cuối: 08/6/2019 https://www.marketsandmarkets.com/Market-Reports/mobility-ondemand-market-198699113.html [19] Techcrunch.com, Uber and Grab hit with $9.5M in fines over ’anti-competitive’ merger Truy cập lần cuối: 08/6/2019 https://techcrunch.com/2018/09/23/uber-and-grab-hit-with-9-5min-fines/ [20] Channelnewsasia.com, Commentary: On both sides of the Causeway, Grab’s grand ambition hits road bumps Truy cập lần cuối: 08/6/2019 https://www.channelnewsasia.com/news/commentary/malaysia- singapore-grab-ride-hailing-ambitions-hit-road-bumps-10995314 [21] Onmilwaukee.com, BidRide, a new local ridershare app, rivals Uber and Lyft Truy cập lần cuối: 08/6/2019 https://onmilwaukee.com/market/articles/bidride-rideshare.html 114 [22] Todayonline.com, ‘Numerous complaints’ about Grab over increased fares and commission fees, says competition watchdog, 2018 Truy cập lần cuối: 08/6/2019 [23] Channelnewsasia.com, Indian ride-hailing firm Jugnoo to enter Singapore with price-bidding system Truy cập lần cuối: 08/6/2019 https://www.channelnewsasia.com/news/singapore/indian-ride- hailing-firm-jugnoo-to-enter-singapore-with-price-10168246 [24] sina.com.cn, Didi, tăng giá lập tức? Didi cuối trả lời (tạm dịch) Truy cập lần cuối: 08/6/2019 http://news.sina.com.cn/o/2017-01-22/doc-ifxzutkf2327804.shtml [25] Trung tâm bồi dưỡng đại biểu dân Ban công tác Đại biểu (Ủy ban Thường vụ Quốc hội), Tổng quan đấu giá hàng hóa thương mại Truy cập lần cuối: 15/7/202 http://tailieu.ttbd.gov.vn:8080/index.php/tai-lieu/tai-lieu-bientap/item/download/510_d1c1f572f262e3e1007933cb90fbde29 115 PHẦN LÝ LỊCH TRÍCH NGANG Họ tên: TRẦN NGỌC BẢO DUY Ngày, tháng, năm sinh: 15/12/1995 Nơi sinh: Bạc Liêu Địa liên lạc: 262/13 Lũy Bán Bích, phường Hịa Thạnh, quận Tân Phú, TP Hồ Chí Minh QUÁ TRÌNH ĐÀO TẠO Tháng 9/2013 – Tháng 4/2018: Sinh viên, chuyên ngành Khoa học Máy tính, trường Đại học Bách Khoa – ĐHQG-HCM Tháng 4/2018 – nay: Học viên cao học, chuyên ngành Khoa học Máy tính, trường Đại học Bách Khoa – ĐHQG-HCM Q TRÌNH CƠNG TÁC Tháng 6/2019 – nay: Kỹ sư hỗ trợ giảng dạy, mơn Khoa học Máy tính, khoa Khoa học Kỹ thuật Máy tính, trường Đại học Bách Khoa – ĐHQG-HCM ... nhận hàng hóa tích hợp hệ thống gọi xe trực tuyến thị trường tiềm trước tiên cần khai thác, hầu hết đơn hàng giao nhận lên lịch trước Việc tích hợp chế đấu giá vào hệ thống giao nhận hàng hóa ứng... cách giải mặt giải thuật cho việc định tuyến cho hệ thống giao nhận hàng hóa dựa chế đấu giá tích hợp vào hệ thống gọi xe trực tuyến, từ đó, góp phần đưa vào thị trường gọi xe cách thức tiếp cận... nghiệp Hàng hóa mang đấu giá khó xác định giá trị thực nên số lượng người tham gia đấu giá thường hạn chế 2.1.3 Ưu điểm đấu giá trực tuyến Đấu giá trực tuyến cách cho phép người tham gia đấu giá