A take-out rider online scheduling method considering fairness

Abstract

This invention discloses an online dispatching method for food delivery riders that considers fairness, including initialization: pre-setting a heterogeneous speed vector of the rider set, a target order acceptance probability vector, and an initial dual variable; dynamic matching decision: when a real-time order arrives, its delivery distance and revenue attributes are extracted, and the marginal benefit of each available rider is calculated through the rider's dual variable, selecting the rider with the maximum benefit for order dispatch; dual variable update: based on the current order dispatch result, the subgradient of the fairness constraint is calculated, and the update step size is determined within the search interval using interpolation, and the dual variable is corrected in real time; iterative closed loop: as the order flow continues to input, the marginal benefit of the rider is repeatedly calculated and the dual variable is updated, and the long-term fairness goal is achieved through feedback adjustment of the dual variable. The dual optimization framework transforms the complex global planning problem into a real-time linear decision, which is suitable for large-scale concurrent scenarios.

Description

Technical Field

[0001] This invention belongs to the field of real-time delivery resource scheduling technology, specifically relating to an online dispatching method for food delivery riders that takes fairness into account. Background Technology

[0002] Rider dispatching strategies are central to platform operations, directly impacting rider earnings and work experience. Existing research on algorithmic fairness largely employs bias-reduction methods based on model building or data processing, primarily constructing multi-objective optimization models for rider workload fairness under offline conditions. However, offline methods have significant drawbacks in practical applications. They presuppose that all order information (input sequence) is known before decision-making, but in reality, orders are highly random and arrive dynamically, making it impossible for the platform to predict future information. In real-world scenarios with large order volumes, frequently recalculating offline optimal solutions based on global information is computationally extremely inefficient and fails to meet the real-time requirements of instant delivery. Existing dispatching methods often treat riders as homogeneous individuals, ignoring their heterogeneity in delivery speed, regional familiarity, etc., leading to a situation where, in pursuing fairness, the overall operational efficiency of the system is often compromised. Summary of the Invention

[0003] This invention proposes a scheduling method that balances delivery efficiency and rider fairness under dynamic online order arrival conditions. This invention fully considers rider heterogeneity and incorporates rider speed... As a core parameter, and based on the "maximum-minimum fairness" criterion, a target number of orders is set for each rider to ensure that scheduling decisions match the rider's actual ability. This invention constructs a Lagrange dual optimization framework, maintaining and dynamically updating a dual variable (i.e., shadow price) reflecting the degree of fairness for each rider. When the dual variable is small, it indicates that the rider's fairness is not fully satisfied, and the system will favor them in subsequent scheduling. By designing an online update algorithm based on interpolation, this invention can use currently observed orders and rider status to correct decisions in real time, achieving a Pareto balance between overall platform benefits and rider fairness without needing to predict future information.

[0004] The technical solution of the present invention is as follows: A method for online dispatching of food delivery riders that considers fairness, comprising the following steps:

[0005] Step 1, Initialization: Preset rider set heterogeneous velocity vector Target order acceptance probability vector and initial dual variables ;

[0006] Step 2, Dynamic Matching Decision: When real-time orders... Upon arrival, the delivery distance and revenue attributes are extracted, and the marginal benefits of each available rider are calculated using the rider's dual variables. The rider with the maximum benefit is selected for order assignment.

[0007] Step 3, Dual Variable Update: Calculate the subgradient of the fairness constraint based on the current order dispatch result, determine the update step size within the search interval using interpolation, and adjust the dual variable in real time. ;

[0008] Step 4, Iterative closed loop: As the order flow continues to input, repeat steps 2 and 3, and achieve the long-term fairness goal through feedback adjustment of the dual variables.

[0009] Preferably, step 1 includes the following steps:

[0010] When the order Upon arrival at the platform, the platform observes the orders. Information and select an available rider , It is a moment Available riders can gather. It's the assembly of all riders, record... As a decision variable, it indicates whether the platform selects a rider. Delivery orders Let vector Indicates time All decisions made;

[0011] when And the platform selects riders Delivery orders hour, Other situations Because of each order Only one rider can be selected for delivery, so for any given time... have , This indicates that there are no available riders to deliver orders. Consequently, the profits from that order will be lost.

[0012] (1).

[0013] Preferably, when a rider is an unavailable rider during the order execution period, then for any given time... and All of them have:

[0014] (2),

[0015] in It is a sufficiently large number, if the order and They were all assigned to the same rider The constraint becomes This represents the execution of the order. Orders must be completed beforehand. If the order and Assigning riders to different riders, this restriction is relaxed to It does not serve a restraining function;

[0016] Each rider has a speed With response rider The ability, when Record the order duration. Let the riders be numbered in descending order of speed, and denote the rider speed vector. .

[0017] Preferably, a target number of orders is set for each rider. And order:

[0018] (3),

[0019] This indicates the riders throughout the entire planning period. Target order quantity It can represent each moment. The platform provides riders The probability of dispatching an order is denoted by a vector. The target order acceptance rate for all riders. The error representing the target number of orders is denoted by a vector. For the target order acceptance error of all riders, note that the inherent meaning of (3) is a target planning, allowing a rider to continue accepting orders after reaching the target number of orders, in order to avoid loss of benefits. Platform decision-makers can design according to specific scenarios. .For example This indicates absolute fairness. This indicates relative fairness based on rider ability. According to the online algorithm of this invention, the number of orders accepted by each rider is almost identical. .

[0020] Preferably, in the food delivery industry, the revenue comes from the user, specifically the difference between the user utility generated by the order and the order delivery time. Each order... It has user utility Users perceive order delivery time to be half the order duration. Let the total benefit over the entire planning period be... ;

[0021] This invention's model focuses on rider fairness. Consider this scenario: at every moment... There are enough riders available to fulfill orders, and rider numbers are arranged in descending order of speed. To maximize efficiency, at any given time... The platform will prioritize selecting the fastest riders. Execute the order, and then at every moment. Platforms will select the fastest riders among available riders, obviously... It is unfair that they will never be assigned a role; therefore, this invention will add a fairness index to the objective function. To avoid similar situations, specifically, this invention defines a concave function. Effect on rider Average number of orders received Ultimately, the value of food delivery services, taking fairness into consideration, is defined as:

[0022] (4),

[0023] Where parameters Used to balance efficiency and fairness The smaller the value, the more the platform focuses on efficiency. The larger the value, the more the platform values ​​fairness.

[0024] Preferably, the offline optimization model solves for a rider scheduling strategy that maximizes efficiency and fairness when all information is known. The optimal solution obtained by the offline optimization model can serve as a performance upper bound, providing a benchmark for evaluating the effectiveness of subsequent online scheduling strategies. Specifically, this includes:

[0025] remember This represents the sequence of information inputs throughout the entire planning period, assuming the sequence of information inputs is known in advance. Based on the objective function and constraints (1), (2), and (3) given in (4) and (6), the offline optimization model is written as follows:

[0026] (7.1),

[0027] (7.2)

[0028] (7.3),

[0029] (7.4).

[0030] Preferably, given an initial dual solution Total number of time periods Target order quantity and step size At any moment According to order information and the current dual solution Calculate the optimal delivery strategy Solve for opportunity cost The constraint (1) must be satisfied. This is a linear and low-dimensional planning problem that can be solved quickly and accurately in most cases. Simultaneously, only currently available riders are assigned, i.e., for... Make an empty decision to satisfy constraint (2), and then process the expected order target. Maximize the fairness option after adjusting for opportunity cost and select By Lemma 2, we know This indicates that the expected order target for each step is a fixed value. Therefore, the algorithm of this invention can help platform decision-makers set targets for all riders in advance through data-driven methods. .

[0031] Compared to existing technologies, the beneficial effects of this invention are as follows: The fairness method driven by rider heterogeneity and the online dispatch method for food delivery riders provided by this invention include the following protected method flow: initializing rider heterogeneity parameters and dual variables, sensing dynamically arriving order characteristics in real time, and updating the dual variables and related step sizes in real time after order dispatch. The protected point includes transforming complex long-term fairness constraints into dynamic dual variables for each rider. In each decision, It acts as a "shadow price" to regulate order dispatch bias: when a rider's historical order acceptance rate is lower than the target, the system automatically lowers their price. This value indirectly enhances the rider's competitiveness in the matching formula for the next order. This transformation reduces global offline optimization to instantaneous online decision-making, with the following specific advantages:

[0032] 1) It can maintain robust decision-making in a real-time environment without needing to predict future order sequences;

[0033] 2) By introducing a velocity variable While ensuring efficiency by allowing those who are capable to do more, it also achieves a fair distribution of workload among riders;

[0034] 3) The dual optimization framework is used to transform the complex global planning problem into a real-time linear decision, which is suitable for large-scale concurrent scenarios. Attached Figure Description

[0035] Figure 1 This is a diagram illustrating the order dispatching model for food delivery services.

[0036] Figure 2 This is a flowchart of the method described in this invention;

[0037] Figure 3 A comparative chart showing the trade-off between fairness and efficiency;

[0038] Figure 4 A trend chart of benefit loss under different fairness weights;

[0039] Figure 5 This is a graph showing the number of orders accepted by riders under the dispatching mode. Detailed Implementation

[0040] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are for illustrative purposes only and are not intended to limit the scope of the invention.

[0041] Example: This example considers an online food delivery platform service with an order dispatch model, such as McDonald's "McDelivery" or Starbucks "Starbucks Delivery". For ease of discussion, such as Figure 1 As shown, consider a restaurant that can meet the order demand within a certain radius around the restaurant. Throughout the planning period, the platform will receive orders sequentially. Sub-order demand, i.e., order At any moment Arrival. Without loss of generality, assuming This is definite and known to the platform. When the platform receives an order, it dispatches a rider from the restaurant to deliver it. After the delivery, the rider returns to the restaurant to wait for the next order. Each order also has its own attributes, including the time when the rider started executing the order (the time the order arrived on the platform). Order duration As a status variable for recording order status, the order's delivery distance Since computers make decisions almost instantaneously, let's assume that the time it takes for each order to arrive at the platform and the time it takes for the rider to receive the assignment and begin executing the order are the same, i.e., both are... Order duration It is the total round-trip time for the rider to complete the order delivery; This is the total round-trip distance for order delivery obtained through route planning. Many methods exist for rapid route planning, so... This is known information to the platform. Furthermore, the platform's service scope is relatively fixed, so it can be assumed that for each... In other words, Independent and identically distributed.

[0042] Furthermore, it includes a rider module:

[0043] When the order Upon arrival at the platform, the platform observes the orders. Information and select an available rider , It is a moment Available riders can gather. It's time for all the riders to assemble. (Note:) As a decision variable, it indicates whether the platform selects a rider. Delivery orders Let vector Indicates time All decisions made. When And the platform selects riders Delivery orders hour, Other situations Because of each order Only one rider can be selected for delivery, so for any given time... have , This indicates that there are no available riders to deliver orders. Consequently, the profits from that order will be lost.

[0044] (1),

[0045] Furthermore, suppose a rider cannot be assigned a new order while fulfilling an order, meaning the rider is unavailable during the order fulfillment period. Then, for any given time... and All of them have:

[0046] (2),

[0047] in It is a sufficiently large number. If the order... and They were all assigned to the same rider The constraint becomes This represents the execution of the order. Orders must be completed beforehand. If the order and Assigning riders to different riders, this restriction is relaxed to It does not serve as a constraint.

[0048] Unlike previous studies that treated riders as homogeneous, this study focuses on the individual attributes of each rider. Each rider has a speed... With response rider The rider's ability is typically calculated from historical data, and its influencing factors include the rider's mode of transportation and their familiarity with the delivery area. Record the order duration. Without loss of generality, let the riders be numbered in descending order of speed, and denote the rider speed vector. .

[0049] Target order quantity:

[0050] From the platform's decision-maker's perspective, to avoid potential risks, he wanted all riders to feel they were being treated fairly. To this end, he set target order completion numbers for each rider. And order:

[0051] (3),

[0052] This indicates the riders throughout the entire planning period. Target order quantity It can represent each moment. The platform provides riders The probability of dispatching an order is denoted by a vector. The target order acceptance probability for all riders. The error representing the target number of orders is denoted by a vector. The target order acceptance error for all riders. Note that the inherent meaning of (3) is a target planning, allowing a rider to continue accepting orders after reaching the target number of orders, in order to avoid loss of benefits. Platform decision-makers can design according to specific scenarios. .For example This indicates absolute fairness. This indicates relative fairness based on rider ability. According to the online algorithm, the number of orders accepted by each rider is almost equal to... .

[0053] Furthermore, in the food delivery industry, profits come from users, specifically the difference between the user utility generated by an order and the delivery time. Each order... It has user utility Users perceive order delivery time to be half the order duration. Let the total benefit over the entire planning period be... .

[0054] The model in this invention focuses on rider fairness. Consider this scenario: at every moment... There are enough riders available to fulfill orders, and rider numbers are arranged in descending order of speed. To maximize efficiency, at any given time... The platform will prioritize selecting the fastest riders. Execute the order, and then at every moment. Platforms will select the fastest riders among available riders, obviously... It's unfair to never be assigned a role. Therefore, a fairness metric will be added to the objective function. To avoid similar situations, specifically, define a concave function. Effect on rider Average number of orders received Ultimately, the value of food delivery services, taking fairness into account, is defined as follows:

[0055] (4),

[0056] Where parameters Used to balance efficiency and fairness The smaller the value, the more the platform focuses on efficiency. The larger the value, the more the platform values ​​fairness.

[0057] Furthermore, it includes an online scheduling strategy module:

[0058] Rider dispatching decisions on food delivery platforms are real-time and dynamic, while order arrivals are highly random, meaning the platform cannot predict future order information when making dispatching decisions. Furthermore, rider availability evolves over time. Therefore, the platform must make decisions at each point in time based on currently observable information, and these decisions cannot be backtracked or reversed. To address this, an online dispatching strategy is proposed.

[0059] Every moment The platform will receive order information. Define the platform's historical information and in accordance with convention Assume the platform is executing a strategy (algorithm). Every moment The platform will base its decisions on the current information. and historical information Make a decision:

[0060] (5),

[0061] For fairness function It adopts the classic Max-Min Fairness.

[0062] (6),

[0063] Minimize the relative number of orders accepted by each rider to ensure that no rider is under-assigned. There are many methods for designing fairness, which will be discussed further in the numerical experiments.

[0064] Before designing the specific online scheduling strategy, we first present an offline optimization model for this problem. The offline optimization model solves for a rider scheduling strategy that maximizes efficiency and fairness, given all available information. The optimal solution obtained from the offline optimization model can serve as an upper bound on performance, providing a benchmark for evaluating the effectiveness of subsequent online scheduling strategies.

[0065] Furthermore, this embodiment includes an offline optimization model:

[0066] For ease of expression, note This represents the sequence of information inputs throughout the entire planning period. When the sequence of information inputs is known in advance... Based on the objective function and constraints (1), (2), and (3) given in (4) and (6), the offline optimization model is written as follows:

[0067] (7.1),

[0068] (7.2)

[0069] (7.3),

[0070] (7.4)

[0071] Furthermore, this embodiment includes Algorithm 1:

[0072] Input: Initial dual solution Total number of time periods Target order , Criterion constant values: , Maximum step size Maximum number of iterations for optimal step size search Maximum number of updates Iteration termination parameter

[0073] Output: Food delivery strategy ;

[0074] for do

[0075] Receive order information

[0076] Make the original decision

[0077] for do

[0078] if do

[0079]

[0080] else do

[0081]

[0082] end

[0083] end

[0084] Execute order strategy

[0085] Calculate a subgradient of the dual objective

[0086] for do

[0087] Initialize step size

[0088] for do

[0089]

[0090] else if Or do

[0091] Calculate using interpolation method

[0092] break

[0093] else if do

[0094]

[0095] break

[0096] else if do

[0097] Calculate using interpolation method

[0098] break

[0099] end if

[0100] else do

[0101]

[0102] end for

[0103]

[0104] if do

[0105] break

[0106] end if

[0107] else do

[0108]

[0109] end for

[0110] end for

[0111] Furthermore, this embodiment includes Algorithm 2 interpolation method. function:

[0112] Input: previous step size and current step size Current gradient Maximum number of iterations Iteration termination parameter

[0113] Output: Optimal step size

[0114] for do

[0115]

[0116] calculate

[0117] if Or do

[0118] Shrink right boundary

[0119] end if

[0120] else if do

[0121]

[0122] break

[0123] else if do

[0124] Update right boundary

[0125] Update left boundary

[0126] else if do

[0127]

[0128] break

[0129] end for

[0130] Furthermore, Algorithm 1 is the core algorithm of this invention. This algorithm is for each rider Maintain and update a dual variable . It's a shadow price for riders, reflecting the degree to which riders are treated fairly: for riders In other words When the size is small, its fairness is not well satisfied; when When the value is large, its fairness is well satisfied. Algorithm 2 is a sub-function interpolation method of Algorithm 1, and its goal is to find the optimal step size through interpolation. Algorithm 2 utilizes the input current step size and the previous step size. Construct a polynomial using the function values ​​and derivatives, and directly search for the minimum points of this polynomial. , The optimal step size.

[0131] During algorithm initialization, an initial dual solution is given. Total number of time periods Target order quantity and step size At that moment According to order information and the current dual solution Calculate the optimal delivery strategy Solving for opportunity cost The constraint (1) must be satisfied. This is a linear and low-dimensional planning problem that can be solved quickly and accurately in most cases. Simultaneously, only currently available riders are assigned, i.e., for... A null decision is made to satisfy constraint (2). Next, the expected order target is processed. Specifically, it is necessary to maximize the fairness item adjusted for opportunity cost and select... However, according to Lemma 2, This indicates that the expected order target for each step is a fixed value. Therefore, the algorithm can help platform decision-makers set targets for all riders in advance using data-driven methods. .

[0132] The method provided by this invention addresses the shortcomings of existing food delivery rider scheduling methods by offering a fairness-considered online food delivery rider scheduling algorithm. Figure 2 As shown. The method runs on the server side and includes the following steps:

[0133] Initialization: Preset rider set heterogeneous velocity vector Target order acceptance probability vector and initial dual variables .

[0134] Dynamic matching decision: When real-time orders Upon arrival, extract its delivery distance and revenue attributes, based on the current dual variables. Set a "shadow price" for riders, calculate the marginal benefit of each available rider, and select the rider with the maximum benefit to assign orders.

[0135] Dual variable update: Calculate the subgradient of fairness constraints based on the current order dispatch result, determine the update step size within the search interval using interpolation, and adjust the dual variable in real time. .

[0136] Iterative closed loop: As the order flow continues to input, steps 2 and 3 are repeated, and the long-term fairness goal is achieved through feedback adjustment of the dual variable.

[0137] Specifically, the performance gap between the proposed online algorithm and the offline optimized benchmark is compared. To this end, a simulated food delivery service scenario is constructed, where the order delivery distance... Independent and identically distributed random variables. Represents the values ​​taken by different fairness weights. The relationship between fairness and efficiency is discussed below. The proposed method will be referred to as the minimax fairness algorithm.

[0138] Figure 3 The horizontal axis represents the fairness function. The value of the ordinate is the benefit function. The value of is shown. The blue line segment represents the solution of offline optimization, and the red line segment represents the solution of online algorithm 1. Compared with offline optimization, the online optimization algorithm does not lose much benefit. This shows that although the online algorithm cannot know future order information in advance when making decisions, it can still approach the offline optimal strategy with complete prior information in terms of benefit, demonstrating the robustness of the algorithm in uncertain environments. Note that with the fairness weight , As the size of the network increases, the improvement in fairness by online algorithms gradually reaches saturation, while system efficiency declines significantly. This phenomenon indicates that excessively large networks... This can lead to algorithms overemphasizing fairness constraints, thereby sacrificing overall operational efficiency. Considering both the improvement in fairness and the loss of benefits, under the simulation settings of this invention, It strikes a relatively balanced balance between the two and is therefore considered a reasonable value.

[0139] The difference in efficiency between online algorithms and offline optimizations is defined as:

[0140] (twenty four),

[0141] Figure 4This indicates that as the planning period T increases, the online optimization algorithm generates... It grows linearly.

[0142] This result is intuitive: as the decision-making cycle lengthens, the locally suboptimal decisions made by the online algorithm at each time step due to incomplete information gradually accumulate, leading to a linear amplification of the overall performance gap over time. However, this growth does not exhibit superlinear or explosive expansion, indicating that the proposed online algorithm still suffers from controllable performance loss in long-term operation.

[0143] A common approach is the heuristic algorithm F-Aware, which identifies riders... Fairness is defined as the ratio of the cumulative value of its currently completed orders to the total value of orders that have arrived in the system, and at each decision point, the rider with the lowest fairness index is given priority to execute new orders.

[0144] The performance of different algorithms was compared under a unified order dispatching model. Specifically, the comparison objects included: (a) a greedy algorithm that only considers maximizing benefits at each time step; (b) the heuristic algorithm F-Aware; and (c) the max-min fairness algorithm, i.e., Algorithm 1. Figure 5 This demonstrates the evolution of rider order acceptance under different algorithms. The three curves, from top to bottom, represent the 0.9 quantile, median, and 0.1 quantile lines after grouping riders according to their abilities, respectively, to characterize the order acceptance distribution characteristics of riders at different ability levels.

[0145] from Figure 5 It can be observed that the maximum-minimum fairness algorithm proposed in this invention can achieve a relatively smooth rider assignment process. The number of orders accepted by riders with different abilities shows a relatively consistent growth trend over time, and the total number of orders accepted by each rider is relatively close at the end of the planning period. In contrast, the F-Aware algorithm fails to achieve smooth assignment of riders during the order assignment process, and its order distribution shows a significant imbalance. Furthermore, there is a large range in the total number of orders accepted by different riders at the end of the planning period. This indicates that a local fairness index constructed solely based on historical value proportions is insufficient to achieve long-term stable fair allocation in dynamic order assignment scenarios.

[0146] Table 1 shows the benefits achieved by the four algorithms. Based on the greedy algorithm that does not consider fairness, it can be found that the maximum-minimum fairness algorithm can achieve better fairness performance without losing too much benefit.

[0147] Table 1 - Algorithm Performance Comparison

[0148]

[0149] The above results collectively demonstrate that using this method helps to achieve a more balanced order allocation in dispatch scenarios with order differentiation, rather than simply meeting fairness requirements at the level of total order volume.

[0150] It should be noted that the above content merely illustrates the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. For those skilled in the art, various improvements and modifications can be made without departing from the principle of the present invention, and all such improvements and modifications fall within the scope of protection of the claims of the present invention.

Claims

1. A fair online dispatching method for food delivery riders, characterized in that, Includes the following steps: Step 1, Initialization: Preset rider set heterogeneous velocity vector Target order acceptance probability vector and initial dual variables ; Step 2, Dynamic Matching Decision: When real-time orders... Upon arrival, the delivery distance and revenue attributes are extracted, and the marginal benefits of each available rider are calculated using the rider's dual variables. The rider with the maximum benefit is selected for order assignment. Step 3, Dual Variable Update: Calculate the subgradient of the fairness constraint based on the current order dispatch result, determine the update step size within the search interval using interpolation, and adjust the dual variable in real time. ; Step 4, Iterative closed loop: As the order flow continues to input, repeat steps 2 and 3, and achieve the long-term fairness goal through feedback adjustment of the dual variables.

2. The method according to claim 1, characterized in that, Step 1 includes the following steps: When the order Upon arrival at the platform, the platform observes the orders. Information and select an available rider , It is a moment Available riders can gather. It's the assembly of all riders, record... As a decision variable, it indicates whether the platform selects a rider. Delivery orders Let vector Indicates time All decisions made; when And the platform selects riders Delivery orders hour, Other situations Because of each order Only one rider can be selected for delivery, so for any given time... have , This indicates that there are no available riders to deliver orders. Consequently, the profits from that order will be lost. （1）。 3. The method according to claim 2, characterized in that, If a rider is not available during the order fulfillment period, then for any given time... and All of them have: （2）， in It is a sufficiently large number, if the order and They were all assigned to the same rider The constraint becomes This represents the execution of the order. Orders must be completed beforehand. ; If order and Assigning riders to different riders, this restriction is relaxed to It does not serve a restraining function; Each rider has a speed With response rider The ability, when Record the order duration. Let the riders be numbered in descending order of speed, and denote the rider speed vector. .

4. The method according to claim 3, characterized in that, Set target order volume for each rider And order: （3）， This indicates the riders throughout the entire planning period. Target order quantity Representing each moment The platform provides riders The probability of dispatching an order is denoted by a vector. The target order acceptance rate for all riders. The error representing the target number of orders is denoted by a vector. To account for the target order acceptance margin for all riders, a rider is allowed to continue accepting orders after reaching the target number of orders. Platform decision-makers design this feature based on specific scenarios. .

5. The method according to claim 4, characterized in that, Each order It has user utility Since order delivery includes round-trip travel, the user's perceived delivery time is set to half the order duration. The total benefit for the entire planning period is denoted as... ; Add a fairness index to the objective function. Define concave functions Effect on rider Average number of orders received Ultimately, the value of food delivery services, taking fairness into consideration, is defined as: （4）， Where parameters It is used to balance efficiency and fairness.

6. The method according to claim 5, characterized in that, The offline optimization model solves for a rider scheduling strategy that maximizes efficiency and fairness when all information is known. The optimal solution obtained by the offline optimization model serves as a performance upper bound, providing a benchmark for evaluating the effectiveness of subsequent online scheduling strategies. Specifically, it includes: remember This represents the sequence of information inputs throughout the entire planning period, assuming the sequence of information inputs is known in advance. Based on the objective function and constraints (1), (2), and (3) given in (4) and (6), the offline optimization model is written as follows: （7.1）， （7.2）， （7.3）， （7.4）。 7. The method according to claim 6, characterized in that, Given an initial dual solution Total number of time periods Target order quantity and step size At any moment According to order information and the current dual solution Calculate the optimal delivery strategy Solve for opportunity cost Constraint (1) must be met, and only currently available riders must be assigned, i.e., for Make an empty decision to satisfy constraint (2), and then process the expected order target. Maximize the fairness option after adjusting for opportunity cost and select .

Images