Keywords

1 Introduction

With the rapid development of the Chinese takeout industry, the number of orders in the peak period has multiplied every year since 2015 [1, 2]. Now, “Mei Tuan”, “Eleme” and “ Star Eleme” basically occupied 90% market share of Chinese takeout industry [1]. The major takeout platforms transferred the core competition from the cultivation of customer consumption habit to the promotion of customer experience gradually. Therefore, Intelligent scheduling system has been vigorously developed by each takeout platform. The statistical data shows that compared with the early manual scheduling, the average delivery time has shortened to 26 min, the average single running distance of riders has reduced by 7%, and the maximum income of rider has increased by 40% since the intelligent scheduling system was deployed [1]. However, some social contradictions are covered up by the beautiful data. As is known to all that it is impossible that the intelligent scheduling system consider the real-world situation well at any time. Some unreasonable scheduling will cause the order delivery delay, however, the riders, who are completely used as tools, have to be punished even if it is not their mistakes. In the long run, it has resulted in the high turnover rate and low happiness index in the takeout industry.

The problem of takeout delivery can be regarded as a kind of expansion problem of Vehicle Routing Problem with Time Windows and Simultaneous Pickup and Delivery problem (VRPPDTW). VRPPD has an important attribute: the vehicle must pick up the goods before delivery; this is obvious in the delivery process. In addition, the location of the rider in the delivery problem is not fixed, namely, the rider will not return to a distribution center after the delivery completed, which is the difference between TDP and general VRPPD problems. Through literature search, it is found that the research on delivery is still limited. Su et al. [3] and Jian et al. [4] have addressed the multi-objective model about the supply chain, coordination mechanism was proved as important to the logistics system. Chen et al. has studied the problem of delivery based on customer satisfaction [5], but it is not enough to only focus on customer satisfaction of time window, and more goals should be considered in the math model. Reyes et al. introduced the meal delivery routing problem firstly, a rolling horizon algorithm has been developed to solve the dynamic meal delivery problem, the computational results showed a nice work of it [6]. LI et al. addressed the delivery problem with an objective function of minimum cost, including the fixed cost, delivery cost and time penalty cost of the new order [7]. However, all of the researches don’t attach the importance of rider’s delivery experiences to the dispatching process. Thus, the bi-objective math model and dispatching method based on Human-Computer Interaction (HCI) have been proposed in this study. In practice, the solution of the TDP would benefit platforms, riders, customers in different aspects: 1) the fuel consumption and total operating cost can be decreased by minimizing the travelled distance and the number of riders employed; 2) the delivery efficiency can be improved enormously 3) the customers satisfaction can be maximized so that the reputation and competitiveness of the platform will be effectively guaranteed; 4) riders will be dispatched with effect and balanced, and their income will increase, most importantly, the riders would find their achievement in the works.

2 Model and Methodology

2.1 Model Construction

This study focuses on two main optimization objectives in the delivery process, namely, cost and satisfaction. Bi-objective function is transformed into single objective optimization by linear weighting method [4]. The objective function is showed in Eq. (1).

$$ \hbox{min} f = a(Z_{11} + Z_{12} + Z_{13} ) - b(w_{1} Z_{21} + w_{2} Z_{22} ) $$
(1)

In Eq. (1), \( a \) and \( b \) are sub-objective function weights. \( Z_{11} \) represents the vehicle operating cost which includes the cost of vehicle depreciation and the salary of the personal. \( Z_{12} \) represents the cost of fuel consumption. \( Z_{13} \) represents the cost of penalty, due to the strict time requirements, a soft time window is used in this study, the order should be delivered within [\( e_{{P_{i} }} ,l_{{P_{i} }} \)], otherwise, additional costs will be incurred and customer satisfaction will be reduced.\( Z_{21} \) represents the customer satisfaction of time. \( Z_{22} \) represents the customer satisfaction of food quality. \( w_{1} \) and \( w_{2} \) are sub-objective function weights. Besides, compared with the traditional VRPPD, most of the constraints of the TDP are similar. While the order merging strategy is considered in the delivery problem, some orders will be merged into one task and then delivered by a rider. Therefore, some constraints are added, that is, an order can only exist in one task, and a task can only be delivered by one rider.

In order to quantify customer satisfaction, and use the algorithm to solve conveniently, the fuzzy evaluation method is used to solve the problem. The relationship between customer satisfaction and membership is shown in Table 1.

Table 1. Relationship between membership value and customer satisfaction
Full size table

2.2 Solution Strategy

In this section, due to the rider’s experiences are considered, the two-stage strategy is introduced to solve the TDP model. As it shown in Fig. 1. In the first stage, insertion detection method and heuristic algorithm are used to solve the initial delivery scheme of the problem. When the rider received the dispatch scheme, the rider can choose to accept the system instruction or some of instructions according to his own distribution experience and current situation. Then, the feedback will be sent back to the system. In the second stage, the problem is solved again by the combination of fixed local solution method and local search algorithm. This scheduling strategy not only makes use of the powerful computing power of the intelligent algorithm, but also combines the personal delivery experience of the rider and real-time environment feedback. Riders with certain autonomy can improve the efficiency and stability of delivery process, more importantly, it will enhance the sense of participation and self-identification of riders and make them enjoy their work finally.

Fig. 1.
figure 1

The HCI solution strategy for takeout delivery problem

Full size image

3 Relevant Algorithms and Technical Means

In order to achieve the above solution strategy, the key technical methods have been fully studied. In this section, all key technical methods will be shown in detail.

3.1 Encoding

The two layers coding method is developed in this study, including the order layer and rider layer. there are 6 orders from 3 o2o restaurants and 3 riders are available. There are 3 orders sent to the first restaurant, 2 orders sent to the second restaurant and 1 order sent to the third takeaway restaurant. Hence, an initial delivery plan is given in Fig. 2. It means the first order and second order from the restaurant 1 are delivered by rider 1 at same time, the first order from restaurant 2 is delivered by rider 2, the third order from restaurant 1 is delivered by rider 3, the second order from restaurant 2 and first order from restaurant 3 are delivered by rider 1 at same time.

Fig. 2.
figure 2

Two-layer coding method

Full size image

3.2 Order Combination Strategy

Around the dinner time, the order placed on the o2o platform would increase sharply. However, due to the limited cost of riders, it is difficult to guarantee that all the orders could be delivered on time unless a balanced schedule is made. Hence, how to merge multiple orders for one rider at a time becomes an important issue of o2o delivery system. Besides the increased delivery efficiency, it also can decrease fuel consumption and food quality consumption. In this model, each order has an urgency attribute, which reflects whether the order is expedited by customers and its duration. The lower the order urgency is, the more urgent the order is. If customers have special requirements for expedited delivery, the order urgency is negative. The most urgent order is placed in group 1 according to its order urgency, and other orders from the same takeaway restaurant are also placed in group 1. Hence, the orders can be sorted according to their order urgency.

Consolidated delivery in this model can greatly improve delivery efficiency, save delivery manpower and decrease fuel consumption. As the orders from same o2o restaurant would be send to different customers, they need to be merged. The combination strategy based on the insertion detection [8] can be seen in Procedure 1.

figure a

3.3 Heuristic Method

A hybrid heuristic algorithm Adaptive Multi-population Genetic Algorithm (AMGA) with two layers coding and adaptive adjustment method to dynamically control crossover rate and mutation rate is developed for order combination and routing planning. Immigrant operator is one of the important operators in AMGA, which can realize the migration operation of excellent chromosomes between different populations and make the optimization speed faster. The immigration operation process is shown in procedure 2.

figure b

4 Simulation Instance Design

Due to the fact that the data of takeout involves the trade secrets of platform competition, there is a lack of standard cases of takeout distribution in the existing instance base. Therefore, combined with the design principles of international standard instance and the service characteristics of Chinese takeout industry, the design of test instances of takeout distribution problem is helpful to test the effectiveness of the proposed model and algorithm. Based on the real environment of Chinese o2o takeout delivery market, we assume that there are 30 customers, 5 restaurants and some riders randomly distributed in a 5-km rectangular range, and the customers send orders to different restaurants at 0. Riders are required to deliver all orders within 40 min. The quantity and type of food ordered by customers should be taken into consideration. Different types of takeout food have different preparation times for the restaurant. The distribution speed of riders can be simplified to uniform driving, but it should be different in peak traffic and flat hump traffic. Customers have a soft time window for food delivery, i.e., overtime is allowed but will be punished. Riders’ service time at each customer should also be considered.

Finally, the solution algorithm can give the following distribution Gantt chart (seen in Fig. 3.). Through this chart, we can clearly know the order consolidation, task allocation, delivery time, rider loading and other data, which will help to identify the problems in the delivery process and formulate improvement plans in the future.

Fig. 3.
figure 3

Gantt chart of the meal delivery

Full size image

5 Conclusions

Based on the brief description of TDP and the comparison with VRPPDTW, this paper puts forward the main characteristics of TDP and constructs the bi-objective mathematical programming model of TDP with the objective of minimum cost and maximum customer satisfaction. In addition, a heuristic genetic algorithm with two-layer coding and adaptive rate of crossover and mutation is proposed. Insertion detection method is used to generate good initial solution. The two-layer coding method is an improvement on traditional VRP coding method, which could solve the problem of order combination and delivery order at the same time. Meanwhile, a Gantt chart is given to indicate the specific delivery sequence of different riders to verify the effectiveness of this proposed method. In further research, a more in-depth and realistic delivery problem might consider a continuous time period to construct a more practical delivery model.