Order scheduling method and device, computer device and computer readable storage medium

Abstract

The application discloses an order scheduling method and device, computer equipment and a computer readable storage medium, relates to the technical field of Internet, dynamically adapts to complex environments such as supply and demand mutation, abnormal weather and the like in actual scheduling, improves the accuracy and efficiency of order scheduling, and has good intelligence. The method comprises the following steps: obtaining a plurality of sample order scheduling schemes, performing multi-dimensional feature index evaluation on each sample order scheduling scheme to obtain a plurality of index scores; through a pre-constructed likelihood estimation function, combining the plurality of index scores corresponding to each sample order scheduling scheme, performing scheme true quality estimation on each sample order scheduling scheme to obtain an implicit quality score; according to the implicit quality score corresponding to each sample order scheduling scheme, collating the plurality of sample order scheduling schemes to obtain a plurality of preference pairs and marking each preference pair; and training an order scheduling model by using the marked plurality of preference pairs and executing a target order scheduling scheme.

Description

Technical Field

[0001] This application relates to the field of Internet technology, and in particular to an order scheduling method, apparatus, computer equipment, and computer-readable storage medium. Background Technology

[0002] In recent years, with the rapid development of internet technology, the order dispatch system, as the core hub connecting user demand and transportation resources in the instant delivery scenario, directly determines the fulfillment efficiency and the experience of all parties involved. Due to the dynamic and random nature of the delivery environment, the order dispatch system needs to process a large number of orders and rider status in real time, and comprehensively consider complex factors such as traffic conditions and weather changes to make optimal order dispatch decisions in a short period of time.

[0003] In related technologies, order scheduling is usually implemented based on rule engines or heuristic algorithms. The order scheduling system will pre-set a series of scheduling rules and weight parameters, such as order distance, rider route compatibility, and estimated completion time. The system will use weighted scoring or greedy strategies to match orders with riders, thereby achieving reasonable order scheduling.

[0004] In the process of developing the relevant technology, the applicant recognized that the relevant technology has at least the following problems: When dealing with complex and dynamic environments, rule-based engines or heuristic algorithms have rigid decision boundaries, making it difficult to cope with unexpected scenarios such as sudden changes in supply and demand or abnormal weather. Furthermore, conflicts are prone to occur between rules, making it difficult to achieve long-term global optimization. This results in low accuracy and efficiency in order scheduling and poor intelligence. Summary of the Invention

[0005] In view of this, this application provides an order scheduling method, apparatus, computer device, and computer-readable storage medium, the main purpose of which is to solve the problem of low accuracy and efficiency in current order scheduling.

[0006] According to a first aspect of this application, an order scheduling method is provided, the method comprising: Multiple sample order scheduling schemes are obtained, and each sample order scheduling scheme is evaluated by multi-dimensional feature indicators to obtain multiple indicator scores; By using a pre-constructed likelihood estimation function and combining multiple index scores corresponding to each sample order scheduling scheme, the true quality of each sample order scheduling scheme is estimated to obtain the implicit quality score corresponding to each sample order scheduling scheme. Based on the implicit quality score corresponding to each of the sample order scheduling schemes, the multiple sample order scheduling schemes are sorted out to obtain multiple preference pairs, and each preference pair is labeled. The order scheduling model is trained using the labeled multiple preference pairs, and when scheduling orders, the order scheduling model is invoked to generate a target order scheduling scheme and execute the target order scheduling scheme.

[0007] According to a second aspect of this application, an order scheduling device is provided, the device comprising: The first evaluation module is used to obtain multiple sample order scheduling schemes and evaluate each sample order scheduling scheme using multi-dimensional feature indicators to obtain multiple indicator scores. The second evaluation module is used to estimate the true quality of each sample order scheduling scheme by combining a pre-constructed likelihood estimation function with multiple index scores corresponding to each sample order scheduling scheme, so as to obtain the implicit quality score corresponding to each sample order scheduling scheme. The preference pair generation module is used to organize the multiple sample order scheduling schemes according to the implicit quality score corresponding to each sample order scheduling scheme, obtain multiple preference pairs, and mark each preference pair; The training module is used to train the order scheduling model using the labeled multiple preference pairs; The scheduling module is used to call the order scheduling model to generate a target order scheduling scheme and execute the target order scheduling scheme when performing order scheduling.

[0008] According to a third aspect of this application, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the method described in any of the first aspects above.

[0009] According to a fourth aspect of this application, a computer-readable storage medium is provided, on which a computer program is stored, wherein the computer program, when executed by a processor, implements the steps of the method described in any one of the first aspects above.

[0010] By employing the aforementioned technical solutions, this application provides an order scheduling method, apparatus, computer device, and computer-readable storage medium. This application acquires multiple sample scheduling schemes and evaluates them using multi-dimensional feature indicators to obtain multiple indicator scores. Then, it utilizes a pre-constructed likelihood estimation function to deeply mine these indicator scores, obtaining implicit quality scores that reflect the true comprehensive quality of the schemes. This achieves automatic extraction of objective, denoised scheduling effect metrics from multi-dimensional indicators, avoiding the subjectivity and rigid boundaries of manually set rule weights. Furthermore, it automatically generates and labels preference pairs based on the implicit quality scores, and uses these preference pairs to train an order scheduling model. This enables the order scheduling model to learn the inherent superiority-inferiority relationships between scheduling schemes from the data, effectively avoiding rule conflicts and local optima problems. Consequently, it dynamically adapts to complex environments such as sudden changes in supply and demand and abnormal weather during actual scheduling, improving the accuracy and efficiency of order scheduling and demonstrating good intelligence.

[0011] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0012] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of this application. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 This paper illustrates a flowchart of an order scheduling method provided in an embodiment of this application. Figure 2 This paper illustrates a flowchart of another order scheduling method provided in an embodiment of this application. Figure 3 This paper shows a schematic diagram of the structure of an order scheduling device provided in an embodiment of this application; Figure 4 A schematic diagram of the device structure of a computer device provided in an embodiment of this application is shown. Detailed Implementation

[0013] Exemplary embodiments of the present application will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present application are shown in the drawings, it should be understood that the present application may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this application will be thorough and complete, and will fully convey the scope of the present application to those skilled in the art.

[0014] This application provides an order scheduling method, such as... Figure 1 As shown, the method includes: S10: Obtain multiple sample order scheduling schemes, evaluate each sample order scheduling scheme using multi-dimensional feature indicators, and obtain multiple indicator scores corresponding to each sample order scheduling scheme.

[0015] The technical solutions in this application embodiment can be applied to order scheduling systems, which can be applied to the field of instant delivery. For example, during peak meal times or in severe weather conditions, delivery platforms face a supply-demand imbalance between a large influx of user orders and limited rider capacity. The order scheduling system can obtain multi-dimensional dynamic data such as the current set of orders to be assigned, the real-time location and load status of online riders, and traffic conditions and weather changes in various delivery areas in real time. It can generate an optimal matching scheme between orders and riders in a short period of time and issue scheduling instructions for execution to balance user experience, rider income and the overall fulfillment efficiency of the platform, ensuring that the delivery system operates stably and efficiently in a high-concurrency and highly random environment.

[0016] In this embodiment, the order scheduling system first extracts multiple sample order scheduling schemes from historical scheduling logs or simulation environments. Each sample order scheduling scheme refers to the set of matching results formed by assigning a specific order to a specific rider at a certain moment. To quantify the initial merits of each scheme, the order scheduling system evaluates each sample order scheduling scheme using multi-dimensional feature indicators. These multi-dimensional feature indicators may include a dispatch interference score determined based on the route compatibility between each assigned order and the original route of the delivery recipient, a wave score determined based on the fulfillment efficiency index of each delivery wave, and a delivery recipient score determined based on the satisfaction or acceptance probability of the delivery recipient for the task combination estimated based on historical data. This constitutes a multi-dimensional feature vector as multiple indicator scores for the sample order scheduling scheme.

[0017] In this way, through the above process, we can initially characterize the surface performance of each sample order scheduling scheme from multiple dimensions, providing basic data for subsequent analysis of its deeper, true quality. For example, we can extract scheduling decision records for each minute of the past week from the historical database, and calculate the average route affinity, average estimated delivery time, average on-time rate, estimated rider satisfaction, and average estimated delivery distance for each order-rider matching scheme in each record. Ultimately, each scheme corresponds to a multi-dimensional feature vector containing five values, which reflects the scheme's indicator scores in each dimension.

[0018] S20: By using a pre-constructed likelihood estimation function and combining multiple index scores corresponding to each sample order scheduling scheme, the true quality of each sample order scheduling scheme is estimated to obtain the implicit quality score corresponding to each sample order scheduling scheme.

[0019] In this embodiment, after obtaining multiple indicator scores for each sample order scheduling scheme, the order scheduling system utilizes a pre-constructed likelihood estimation function to perform in-depth analysis of these multiple indicator scores. The likelihood estimation function is constructed based on a latent variable model, with the core assumption that each sample order scheduling scheme possesses an unobservable latent quality score representing the scheme's true overall quality. The multiple indicator scores across multiple dimensions are observations of this latent quality score mapped through a nonlinear response function and superimposed with Gaussian noise. The nonlinear response function characterizes the variation of different dimensional indicators with the true quality, while the variance of the Gaussian noise reflects the reliability of the observations in each dimension. Thus, by maximizing the likelihood estimation function and simultaneously optimizing the latent quality score, nonlinear response function parameters for each dimension, and noise variance for each sample order scheduling scheme, a better fit to the observed multidimensional indicators can be achieved. Furthermore, once the optimization converges, the optimized solution corresponding to each sample order scheduling scheme is the latent quality score of that scheme.

[0020] Through the above process, objective and denoised scheduling performance metrics can be automatically inferred from noisy and multi-source service indicators, overcoming the subjectivity and rigid boundaries of traditional methods that rely on manually setting indicator weights. For example, the five-dimensional feature vector of each sample order scheduling scheme is used as observation data and input into a latent variable model for maximum likelihood estimation. After iterative optimization, a numerical latent quality score is obtained for each scheme, which comprehensively reflects the true level of the scheme in terms of efficiency, user experience, etc.

[0021] S30: Based on the implicit quality score corresponding to each sample order scheduling scheme, organize the multiple sample order scheduling schemes to obtain multiple preference pairs and label each preference pair.

[0022] In this embodiment, after obtaining the implicit quality scores of all sample order scheduling schemes, the order scheduling system compares and organizes the sample order scheduling schemes pairwise based on these scores to generate multiple preference pairs. For example, assuming a scheme... The implicit quality score is higher than the proposed solution. The order dispatch system will then determine the solution. It outperforms the solution in terms of actual quality. The two options are combined into a preference pair, and their superiority / inferiority relationship is labeled as "". Superior In this way, by traversing all sample order scheduling schemes, a large number of labeled preference pairs can be generated, thereby realizing the automatic generation of preference labels without manual intervention or pre-setting complex rules, ensuring the consistency and objectivity of training data, and providing reliable supervision signals for subsequent learning of the inherent superiority and inferiority relationships between scheduling schemes.

[0023] For example, assuming that the implicit score of sample order scheduling scheme A is 8.5, the implicit score of sample order scheduling scheme B is 7.2, and the implicit score of sample order scheduling scheme C is 9.0, then preference pairs (A is better than B), (C is better than A), and (C is better than B) are generated, and each preference pair is assigned a corresponding superiority or inferiority label.

[0024] S40: Train the order scheduling model using multiple labeled preference pairs, and when scheduling orders, call the order scheduling model to generate a target order scheduling scheme and execute the target order scheduling scheme.

[0025] In this embodiment, the order scheduling system trains an order scheduling model using generated labeled preference pairs. The order scheduling model takes the features of the order scheduling scheme as input and outputs a scheme score. Specifically, during model training, for each labeled "..." Superior Based on the preferences, the order scheduling model will calculate a solution. With the plan The score difference is calculated and mapped to the scheme using a normalized exponential function. Superior Solution The predicted probability is then used to construct a negative log-likelihood loss function. By minimizing this loss, the model parameters are optimized so that the scheme score output by the order scheduling model can accurately reflect the order of true quality.

[0026] After model training, in actual order scheduling scenarios, when scheduling decisions are needed, the order scheduling system acquires real-time data to construct multiple candidate scheduling schemes. The system then scores these schemes, selecting the highest-scoring scheme as the target order scheduling scheme and executing the corresponding order allocation operation. In this way, the order scheduling model can automatically learn the inherent advantages and disadvantages between sample order scheduling schemes from the data, effectively avoiding rule conflicts and local optima problems. It can also dynamically adapt to complex environments such as sudden changes in supply and demand and abnormal weather, thus significantly improving the accuracy, efficiency, and intelligence of online scheduling. For example, a large number of generated preference pairs can be used as training data to train an order scheduling model. In actual operation, facing 10 orders and 20 riders to be assigned, multiple candidate matching schemes are generated in real time and scored. Finally, the scheme with the highest score is selected to assign the order to the corresponding rider to complete the scheduling.

[0027] Optionally, in this embodiment of the application, the step of evaluating each sample order scheduling scheme with multi-dimensional feature indicators to obtain multiple indicator scores corresponding to each sample order scheduling scheme includes: for each sample order scheduling scheme, evaluating the sample order scheduling scheme with feature indicators in multiple dimensions to obtain multiple indicator scores, wherein the multiple indicator scores include one or more of dispatch interference scores, wave scores, and delivery object scores.

[0028] In this embodiment, when the order scheduling system evaluates each sample order scheduling scheme using multi-dimensional feature indicators, it conducts quantitative analysis from three core dimensions to construct an indicator scoring system that comprehensively reflects the overall performance of the scheduling scheme. The first dimension is the dispatch interference score, which measures the degree of disturbance caused by a single dispatch action to the existing task planning of the delivery recipient. Specifically, for each assigned order in the sample order scheduling scheme, the order scheduling system obtains the new path planning node sequence of the delivery recipient wave after the order is assigned. By calculating the deviation between the pick-up and delivery points of the assigned order and the paths of other orders within the wave, it obtains the order's routeability index. A higher routeability index indicates less interference with the original route of the delivery recipient. Subsequently, the average routeability index of all assigned orders in the sample order scheduling scheme is taken to obtain the dispatch interference score of the scheme. For example, in a historical dispatch record, the rider originally planned to deliver two orders, A and B, in sequence. At this time, a new order, C, was assigned. After the order dispatch system replanned the route, it found that the pick-up point of order C was located on the original route, and the delivery point was also near order B. Therefore, the route of order C was relatively convenient, and the corresponding order dispatch interference score was also relatively high, indicating that this order dispatch was relatively friendly to the rider.

[0029] The second dimension is the wave score, which measures the overall efficiency and fulfillment reliability of the scheduling scheme. For each rider wave formed in the sample order scheduling scheme, a rider wave refers to a list of order tasks currently undertaken by the rider. The order scheduling system calculates the core efficiency indicators for that wave, specifically including time-related indicators such as the estimated total delivery time of the wave, distance-related indicators such as the average delivery distance of orders within the wave, and fulfillment-related indicators such as the on-time delivery probability of the wave predicted based on historical data. Subsequently, the values ​​of all waves on the above indicators are averaged to obtain the wave score of the corresponding sample order scheduling scheme. Taking the scheduling during the lunch peak as an example, a wave contains three orders. The order scheduling system estimates its total delivery time to be 45 minutes, the average delivery distance to be 2.5 kilometers, and the on-time rate to be 92%. These values ​​together constitute a quantitative description of the efficiency and reliability of the wave. A high score means that the wave has achieved better timeliness and stability globally.

[0030] The third dimension is the delivery recipient rating, specifically reflected in rider evaluations, which aims to ensure the health of the delivery capacity ecosystem and the rider experience. For each rider wave in the sample order dispatch scheme, the order dispatch system estimates the potential subjective rider evaluations for that wave based on historical data of rider feedback patterns for different task combinations. This includes predicting the probability of a rider accepting the task combination or their satisfaction level after completing the task. Then, the estimated rider feedback from all waves is averaged to obtain the rider evaluation score for the scheme, which is also the delivery recipient rating. For example, if a wave's task combination allows a rider to complete all orders along their route without exceeding their expected working hours, the estimated rider evaluation score will be higher; conversely, if the wave includes orders from remote areas or causes riders to take significant detours, the score will be correspondingly lower.

[0031] After obtaining the scores of the above three dimensions, the order scheduling system uses these scores as multiple scores for the sample order scheduling scheme. This allows for a complete characterization of the sample order scheduling scheme's real-time performance from three levels: micro-interference, global efficiency, and capacity ecosystem. This provides multi-dimensional basic observation data for subsequent in-depth analysis of the scheme's true quality, effectively avoiding evaluation bias caused by a single indicator and laying a solid data foundation for subsequent hidden value modeling.

[0032] Optionally, in this embodiment, the sample order scheduling scheme is evaluated using feature indicators across multiple dimensions to obtain multiple indicator scores, including: determining each sample delivery object to be dispatched according to the sample order scheduling scheme; dispatching the specified orders to each sample delivery object according to the sample order scheduling scheme and the currently accepted orders of each sample delivery object; generating an updated delivery path for each sample delivery object; querying the current delivery path of each sample delivery object; and calculating the route compatibility between the updated delivery path and the current delivery path of each sample delivery object to obtain a dispatch interference score; and / or, dispatching the specified orders to each sample delivery object according to the sample order scheduling scheme. For each sample delivery object, a list of delivery tasks is generated after the sample order scheduling plan is completed. Based on the delivery task list, a global efficiency index is calculated for each sample delivery object, and the average global efficiency index of all sample delivery objects is calculated to obtain a wave score. Alternatively, the feedback data on the delivery task list corresponding to each sample delivery object after the sample order scheduling plan is completed is estimated to obtain an estimated feedback value for each sample delivery object, and the average estimated feedback value of all sample delivery objects is calculated to obtain a delivery object score.

[0033] In this embodiment, when the order scheduling system evaluates the sample order scheduling scheme using multi-dimensional feature indicators to obtain multiple indicator scores, the calculation is carried out from three dimensions respectively: The first dimension is the calculation of the dispatch interference score. Specifically, firstly, the order scheduling system identifies all sample delivery objects involved in the sample order scheduling plan. These sample delivery objects can be riders, intelligent delivery robots, intelligent delivery vehicles, etc. For each sample delivery object, the sample order scheduling plan specifies one or more orders to be assigned to that object. Then, based on the sample delivery object's currently accepted orders (i.e., the unfinished orders it has at the time of scheduling) and the newly assigned orders, the system replans the delivery route for the sample delivery object, resulting in an updated delivery route. Simultaneously, it queries the sample delivery object's current delivery route without accepting new assignments, i.e., the optimal route containing only the original orders. By comparing the updated delivery route with the current delivery route, the system calculates the route affinity caused by the new order being taken due to detours or changes in order. This affinity can be quantified, for example, by the difference in the total time or total distance between the two routes. A higher route affinity indicates less interference from the new order on the original delivery object's plan. Finally, the average route affinity of all delivery objects is taken to obtain the dispatch interference score for the sample order scheduling plan. For example, a rider originally had two orders, A and B. The optimal route was to deliver A first and then B. After a new order C was assigned, the rider replanned the route and found that picking up C first and then delivering C could be conveniently inserted between A and B, making the detour distance very short. In this case, the rider's route convenience score was high, contributing a high order dispatch interference score.

[0034] The second dimension is the calculation of the wave score. Specifically, for each sample delivery object in the sample order scheduling scheme, the order scheduling system generates a delivery task list for that sample delivery object after the sample order scheduling scheme is completed, based on the designated orders assigned to that sample delivery object according to the sample order scheduling scheme and the orders that the sample delivery object has already accepted. This delivery task list is also known as the wave. Then, for the delivery task list of each sample delivery object, a global efficiency index is calculated. The global efficiency index can specifically include time-dimensional indicators such as the estimated total delivery time of the wave, distance-dimensional indicators such as the average delivery distance of orders within the wave, and fulfillment-dimensional indicators such as the probability of on-time delivery of the wave predicted based on historical data. After calculating each index for each sample delivery object separately, the average of the indices for all sample delivery objects is calculated to obtain the comprehensive performance of the sample order scheduling scheme in terms of global efficiency, i.e., the wave score. For example, during the midday peak dispatch, a sample order dispatch plan assigns three orders to rider A, with an estimated total time of 50 minutes, an average distance of 2 kilometers, and an on-time rate of 90%; rider B is assigned two orders, with a total time of 30 minutes, an average distance of 1.5 kilometers, and an on-time rate of 95%. The average of all indicators for all riders is the wave score of this sample order dispatch plan. A high score means that the sample order dispatch plan has achieved high delivery efficiency and reliability overall.

[0035] The third dimension is the calculation of the delivery recipient score. Specifically, the order dispatch system focuses on the health of the delivery capacity ecosystem and rider experience, thus introducing the delivery recipient score dimension. For each sample delivery recipient in the sample order dispatch plan, the order dispatch system estimates the possible feedback value of the sample delivery recipient to this task list based on the corresponding delivery task list after the sample order dispatch plan is completed, combined with historical data on the feedback patterns of sample delivery recipients to different task combinations, such as acceptance rate, complaint rate, and satisfaction surveys. This feedback value can be quantified into a numerical score to represent the rider's satisfaction or willingness to accept. Subsequently, the estimated feedback values ​​of all sample delivery recipients are averaged to obtain the delivery recipient score for the sample order dispatch plan. For example, if a rider's task list contains three orders, the route planning is reasonable and does not exceed their preferred working hours, then their estimated feedback value is high; conversely, if the task list contains a large number of remote orders or causes the rider to drive empty for a long time, the feedback value is low.

[0036] In this way, through the separate calculations of the above three dimensions, the order scheduling system can fully characterize the real-time performance of the sample order scheduling scheme from three levels: micro-interference, global efficiency, and delivery recipient experience. These index scores serve as the basic observation data for subsequent implicit value inference, effectively avoiding the one-sidedness of single index evaluation, and providing rich and objective feature inputs for training a high-quality order scheduling model.

[0037] In this embodiment of the application, optionally, before estimating the true quality of each sample order scheduling scheme by combining a pre-constructed likelihood estimation function with multiple indicator scores corresponding to each sample order scheduling scheme to obtain the implicit quality score corresponding to each sample order scheduling scheme, the method further includes: determining multiple preset service indicators, and establishing an observation model by combining multiple sample order scheduling schemes and multiple service indicators. In the observation model, each sample order scheduling scheme corresponds to an unobservable implicit quality variable, and each service indicator is represented as the observation result of the corresponding implicit quality variable after being mapped by a nonlinear response function and superimposed with Gaussian noise. The variance of the Gaussian noise is used to characterize the observation reliability of the corresponding service indicator. Based on the observation model, for multiple sample order scheduling schemes and multiple service indicators, a log-likelihood function is constructed as a likelihood estimation function, using the implicit quality score of each sample order scheduling scheme, the nonlinear response function parameter of each service indicator, and the noise variance of each service indicator as variables. The likelihood estimation function is used to quantify the joint probability distribution of the actual values ​​of each service indicator of each sample order scheduling scheme observed under the values ​​of the variables.

[0038] In this embodiment, to infer the true quality of each sample order scheduling scheme from observable indicators across multiple dimensions, a mathematical model capable of characterizing the observation data generation process, i.e., a pre-constructed likelihood estimation function, is first established. Specifically, when constructing the likelihood estimation function, several preset service indicators are first determined. These service indicators correspond to the various dimensions of the multiple indicator scores calculated in the preceding steps, such as the route proximity score for dispatch interference, the estimated delivery time and on-time rate for wave score, and the rider feedback estimate for delivery recipient score. For multiple sample order scheduling schemes and these service indicators, the order scheduling system constructs an observation model. The core assumption of the observation model is that each sample order scheduling scheme possesses an unobservable latent quality variable. This variable represents the true overall quality of the solution, and is the essential attribute hidden beneath multiple indicators; while each observable service indicator ,in Indicates the first Each service metric is the result of superimposing random noise after the implicit quality variable has undergone a specific mapping.

[0039] The likelihood estimation function can be expressed by the mathematical formula shown in Formula 1 below: Formula 1:

[0040] In Formula 1, For the first The nonlinear response function corresponding to each service metric is a nonlinear function used to characterize the actual overall quality of the scheduling scheme. The changes in the service indicator should theoretically reflect how it should change accordingly. For example, as scheduling quality improves, the on-time performance indicator may show a non-linear growth trend of first rising rapidly and then saturating. To observe the noise, it follows a pattern with a mean of 0 and a variance of . The Gaussian distribution of the noise term reflects the random fluctuations in the actual service performance observation process. For example, the same scheduling quality may have slightly different on-time rates due to occasional road condition fluctuations; while the variance... The magnitude of the variance directly reflects the observation reliability of the corresponding service indicator. The smaller the variance, the more stable and reliable the indicator is, and the greater its contribution to the inference of implicit quality should be.

[0041] In this embodiment, optionally, the true quality of each sample order scheduling scheme is estimated by combining a pre-constructed likelihood estimation function with multiple indicator scores corresponding to each sample order scheduling scheme, to obtain the implicit quality score corresponding to each sample order scheduling scheme. This includes: optimizing the implicit quality score of each sample order scheduling scheme, the nonlinear response function parameters of each service indicator, and the noise variance of each service indicator by maximizing the likelihood estimation function and combining multiple indicator scores corresponding to each sample order scheduling scheme, so that the fit of the likelihood estimation function to the actual observed data reaches a preset convergence condition; and outputting the implicit quality score of each sample order scheduling scheme obtained after optimization and convergence as the implicit quality score corresponding to each sample order scheduling scheme.

[0042] In this embodiment, a true quality estimate is performed on each sample order scheduling scheme using a pre-constructed likelihood estimation function to obtain an implicit quality score. The core of this approach lies in performing an optimization solution process for maximum likelihood estimation. Specifically, the order scheduling system will provide the acquired... Individual sample order scheduling schemes and each scheme in The index matrix, composed of the actual observations of each service indicator (i.e., the vector calculated in the previous steps to indicate the scores of multiple indicators), is solved by maximizing the previously constructed log-likelihood function and simultaneously optimizing three types of unknown parameters: the first type being the implicit quality score of each sample order scheduling scheme. ,in The value ranges from 1 to This is the ultimate goal to be output; the second category is the non-linear response function corresponding to each service metric. The parameters, where The value ranges from 1 to These functions characterize the nonlinear relationship between different service metrics and the actual scheduling quality; the third type is the noise variance of each service metric. It is used to characterize the reliability of the observed values ​​of each service indicator.

[0043] The objective function of the above optimization process is the mathematical expression for the maximum likelihood estimation shown in Equation 2 below: Formula 2:

[0044] In Formula 2, Indicates the first The implicit quality score of each sample order scheduling scheme is a quantitative representation of the scheme's true overall quality. Indicates the first The nonlinear response function corresponding to each service indicator maps implicit quality to the theoretical predicted value of that indicator. For the first The noise variance of a service indicator reflects the degree of fluctuation in the observed values ​​of that indicator; It is the first The sample at the th The actual observed values ​​of each service indicator, that is, the values ​​of the multiple indicator scores calculated in the previous steps on the corresponding dimensions. The first term in Formula 2... The second term represents the penalty for prediction error, the magnitude of which is determined by the squared error and the noise variance: for indices with smaller noise variance, the same prediction error will result in a larger penalty, thus guiding the optimization process to place greater emphasis on accurate fitting of these indices; This constrains the noise variance itself, preventing it from increasing indefinitely and causing model failure. By summing over all samples and all indicators, Equation 2 comprehensively measures the joint probability of generating all observed data under the current parameter settings.

[0045] In the actual solution process, the order scheduling system employs iterative optimization algorithms such as gradient descent or the EM algorithm to maximize the likelihood estimation function, gradually improving its fit to the actual observed data until a preset convergence condition is met, such as parameter changes being less than a threshold or the number of iterations reaching an upper limit. When the optimization converges, the optimization variables corresponding to each sample order scheduling scheme are... This is the implicit quality score of the scheme, which aims to inversely infer the true quality from multiple noisy observation indicators.

[0046] Thus, the implicit quality score of each solution is obtained in the end. This is a single, denoised numerical value that integrates multi-dimensional information. It reflects the true quality level of the scheduling scheme more stably than any single original indicator, providing an objective basis for generating reliable preference labels. Through this implicit value inference method based on maximum likelihood estimation, the embodiments of this application can automatically learn the importance weights of different service indicators and their variation with quality, thereby eliminating the subjectivity of manually setting indicator weights in traditional methods and achieving an objective measurement of the true quality of the scheduling scheme from a data-driven perspective.

[0047] Optionally, based on the implicit quality score corresponding to each sample order scheduling scheme, multiple sample order scheduling schemes are organized to obtain multiple preference pairs, and each preference pair is marked. This includes: comparing the implicit quality scores corresponding to every two sample order scheduling schemes; combining two sample order scheduling schemes with a relationship between their implicit quality scores into a preference pair to obtain multiple preference pairs; identifying the relationship between the implicit quality scores corresponding to the two sample order scheduling schemes in each preference pair; and marking each preference pair.

[0048] In this embodiment, the order scheduling system generates preference labels based on the implicit quality score corresponding to each sample order scheduling scheme, thereby achieving automated data labeling based on numerical comparison. Specifically, the implicit quality score of each sample order scheduling scheme is obtained through maximum likelihood estimation. Subsequently, these scores quantified the true overall quality of each scheduling scheme in numerical form. Based on this, a systematic analysis of multiple sample order scheduling schemes was conducted: first, the implicit quality scores of every two sample order scheduling schemes were compared, for example, scheme... and plan If a solution is found Implicit quality score Larger than the plan rating If a clear superiority-inferiority relationship exists between the two options, they are combined into a preference pair, and the relative importance of these two options is identified, assigning a corresponding label to the preference pair. By traversing all possible combinations of options, or by employing sampling strategies such as random sampling or filtering based on rating differences, a finite number of preference pairs can be generated, thus obtaining a dataset containing a large number of labeled preference pairs.

[0049] In this way, through the above process, without any human intervention or the need to preset complex rules or weight parameters, relying solely on the implicit quality score objectively inferred from multidimensional indicators, the noise introduced by traditional methods due to inconsistent manual labeling, rule conflicts, or subjective setting of indicator weights is effectively avoided, ensuring the inherent consistency and objectivity of the training data, and laying a reliable foundation for the subsequent construction of an intelligent model that can accurately understand the true quality of the scheduling scheme.

[0050] For example, suppose the implicit quality scores of three historical scheduling schemes are obtained through the aforementioned steps: Scheme A indicates global scheduling during the lunch peak period, with a score of 8.5; Scheme B indicates local area scheduling during the same period, with a score of 7.2; and Scheme C indicates emergency scheduling under severe weather conditions, with a score of 9.0. The order scheduling system automatically compares these schemes and generates preference pairs: (A is better than B), (C is better than A), and (C is better than B), and records the corresponding superiority / inferiority label for each preference pair. In this way, even if there are a large number of sample order scheduling schemes in actual practice, such as thousands of historical scheduling records, the order scheduling system can complete all pairwise comparisons in a short time and generate a large number of labeled preference pairs for subsequent model training.

[0051] In this embodiment of the application, optionally, the implicit quality scores corresponding to the two sample order scheduling schemes in each preference pair are identified, and each preference pair is labeled, including: for each preference pair, one sample order scheduling scheme in the preference pair is designated as the first sample order scheduling scheme, and the other sample order scheduling scheme in the preference pair other than the first sample order scheduling scheme is designated as the second sample order scheduling scheme, wherein the implicit quality score corresponding to the first sample order scheduling scheme is greater than the implicit quality score corresponding to the second sample order scheduling scheme; a label indicating that the first sample order scheduling scheme is superior to the second sample order scheduling scheme is generated, and the preference pair is labeled using the label.

[0052] In this embodiment, for each preference pair determined through pairwise comparisons, the order scheduling system first identifies the implicit quality scores of the two sample order scheduling schemes in the preference pair, designating the scheme with the higher score as the first sample order scheduling scheme and the scheme with the lower score as the second sample order scheduling scheme. Subsequently, a label is generated to indicate the superiority relationship. This label explicitly expresses the preference information that "the first sample order scheduling scheme is superior to the second sample order scheduling scheme," and the label is attached to the corresponding preference pair to complete the labeling process. This transforms the abstract numerical differences into training signals with clear semantics, enabling the subsequent order scheduling model to clearly understand the superiority indication conveyed by each preference pair.

[0053] For example, suppose implicit quality scores for three historical scheduling schemes are obtained through implicit value estimation. Scheme P, which uses a route-first strategy during the midday peak, has a score of 9.2; Scheme Q, which uses a distance-first strategy during the same time period, has a score of 7.8; and Scheme R, which uses a balanced strategy, has a score of 8.5. When processing preference pairs, the order scheduling system generates multiple preference pairs such as (P,Q) and (P,R). For the preference pair (P,Q), since 9.2 > 7.8, the order scheduling system labels scheme P as the first sample order scheduling scheme and scheme Q as the second sample order scheduling scheme, generating the label "Scheme P is superior to Scheme Q". Similarly, for the preference pair (P,R), scheme P is designated as the first scheme and scheme R as the second scheme, generating the label "Scheme P is superior to Scheme R". In this way, each preference pair obtains a clear and consistent indication of superiority or inferiority, avoiding subjective ambiguity that may arise from manual labeling.

[0054] The method of role classification and label generation based on objective numerical values ​​described above ensures the internal consistency and scalability of the preference dataset. In large-scale datasets containing thousands of samples, the order scheduling system can process all preference pairs with uniform rules. Regardless of the sample size, it can guarantee strict consistency of the labeling logic and avoid fatigue errors or standard drift common in manual labeling. Furthermore, by clearly distinguishing the roles of the first and second schemes, a standardized input format is provided for subsequent preference learning. This allows the model to accurately learn the inherent rules of "why a good scheduling scheme is better than a bad one," thereby improving the discriminative ability and generalization performance of the final order scheduling model. This eliminates the reliance on manually set rule weights, making the entire training data preparation process more objective and efficient, and laying a solid data foundation for building an intelligent order scheduling system.

[0055] Optionally, in this embodiment, an order scheduling model is trained using multiple labeled preference pairs, including: constructing an order scheduling model to be trained; calculating the difference between the implicit quality scores corresponding to the two sample order scheduling schemes in each preference pair, inputting each labeled preference pair and its corresponding difference into a preset normalized exponential function to obtain the predicted probability corresponding to each preference pair, wherein the predicted probability is used to indicate the probability that the sample order scheduling scheme with a high implicit quality score in the corresponding preference pair is better than the sample order scheduling scheme with a low implicit quality score; constructing a negative log-likelihood loss function based on the predicted probabilities corresponding to multiple preference pairs, wherein the negative log-likelihood loss function is used to measure the difference between the actual superiority / inferiority relationship and the predicted probability in each preference pair; iteratively updating the parameters of the order scheduling model with minimizing the negative log-likelihood loss function as the optimization objective, so that the scheme score output by the order scheduling model converges, thereby obtaining the trained order scheduling model.

[0056] In this embodiment, the order scheduling system trains an order scheduling model using multiple labeled preference pairs. During model training, the Bradley-Terry preference learning framework is introduced to transform preference labeling into a probabilistic modeling problem. Optimization is then performed to ensure that the model's output score accurately reflects the true merits of different scheduling options. Specifically, a training order scheduling model is first constructed. This model can be a deep neural network or other parameterized functions. Its input consists of multidimensional features of the sample order scheduling options, such as route proximity, estimated time, and rider evaluation metrics. The output is a scalar option score. This is used to quantify the merits of the scheduling scheme.

[0057] For each labeled preference pair, the preference pair contains two sample order scheduling schemes, assuming the schemes are... Marked as superior to the solution That is, the plan The implicit quality score is higher than the proposed solution. In the Bradley-Terry model, the solution... Superior Solution The probability can be defined using the expression shown in Formula 3 below: Formula 3:

[0058] In formula 3, and These represent the current order scheduling model's response to the proposed solution. and plan The output score of the solution; This is an exponential function used to convert scores into non-negative weights; This represents the normalization exponential function, specifically the sigmoid function, whose mathematical form is: It can convert the input real number difference Mapping to the (0, 1) interval, the output value is the solution. Superior Solution The predicted probability. That is, Formula 3 indicates that the probability of one option being superior to the other is determined by the difference in their scores via a sigmoid function; the larger the score difference, the better the outcome of the two options. Superior Solution The closer the probability is to 1, the closer it is to 0.5; conversely, if the difference in scores is smaller or even negative, the probability approaches 0.5 or lower.

[0059] Based on the predicted probabilities calculated for each preference pair, the order scheduling system constructs a negative log-likelihood loss function. The optimization objective of the negative log-likelihood loss function is to minimize its value, that is, to minimize the value of each preference pair when the model's output score accurately reflects the true superiority-inferiority relationship. It is positive and relatively large. When the value is close to 1, the corresponding negative logarithmic loss is close to 0; conversely, if the model output contradicts the true label, the loss value will increase significantly, resulting in a large gradient update during the optimization process.

[0060] This preference learning method based on the Bradley-Terry model enables the order scheduling model to automatically learn the relationship between the features of scheduling schemes and their final merits from a large amount of pairwise comparison data, without the need for manually designing complex scoring rules or weight parameters. After training convergence, the order scheduling model has the ability to score any new scheduling scheme; a higher score indicates a better scheme. Thus, in actual online scheduling, multiple candidate scheduling schemes at the current time are simply input into the order scheduling model, and the output scores are compared to quickly select the better scheme. This effectively overcomes the rigidity of traditional rule engines in complex environments and can adaptively mine the nonlinear impact of different features on scheduling performance from the data, thereby improving the overall accuracy and intelligence of order scheduling.

[0061] Optionally, in this embodiment of the application, when performing order scheduling, the order scheduling model is invoked to generate a target order scheduling plan and execute the target order scheduling plan, including: when performing order scheduling, obtaining real-time scheduling data at the current moment, the real-time scheduling data including information on orders to be assigned, information on available delivery objects, and environmental status information; referring to the real-time scheduling data, constructing multiple candidate order scheduling plans, and inputting the multiple candidate order scheduling plans into the order scheduling model, so that the order scheduling model calculates a plan score for each candidate order scheduling plan and outputs it; based on the plan score of each candidate order scheduling plan output by the order scheduling model, comparing the scores of the multiple candidate order scheduling plans, and selecting the candidate order scheduling plan with the highest plan score as the target order scheduling plan; and performing an order allocation operation according to the target order scheduling plan, assigning the corresponding orders to the corresponding delivery objects.

[0062] In this embodiment, after completing the offline training of the order scheduling model, the order scheduling system can execute real-time order allocation tasks through the order scheduling model. Specifically, when the order scheduling system needs to perform order scheduling, it first obtains the real-time scheduling data at the current moment. The real-time scheduling data can include information on orders to be allocated, information on available delivery objects, and environmental status information. Among them, the information on orders to be allocated can include information such as the pickup location, delivery location, order time, and estimated preparation time for each order; the information on available delivery objects can include riders who are currently idle or about to complete the previous order, including their real-time location, current load, delivery routes of accepted orders, and working hours; the environmental status information can include information such as real-time traffic conditions, weather conditions, and time period characteristics of each delivery area. These data together depict the complete supply and demand situation and environmental background at the current scheduling moment, providing a basis for generating a reasonable scheduling plan.

[0063] After acquiring real-time scheduling data, the order scheduling system constructs multiple candidate order scheduling schemes based on this data. The generation of candidate order scheduling schemes can employ heuristic search or sampling methods. For example, based on the current set of orders to be assigned and the set of available riders, it enumerates some reasonable order-rider matching combinations to form several possible allocation results. Each result is a candidate order scheduling scheme, and each candidate order scheduling scheme corresponds to a specific set of assignment relationships, such as which orders are assigned to which riders and how each rider's task list changes after receiving a new order. Subsequently, all candidate order scheduling schemes are input into a pre-trained order scheduling model. Having undergone the aforementioned preference learning training, the order scheduling model is capable of scoring any scheduling scheme. The model takes the multi-dimensional features of each candidate scheme as input, calculates a scalar scheme score through forward propagation, and outputs it. This score represents the quantitative evaluation of the scheme's merits by the order scheduling model; a higher score indicates a better scheme. For example, suppose there are currently 8 orders to be assigned, and 5 available riders in the system, each with different locations and load statuses. Based on real-time data, the order scheduling system generates 20 candidate order scheduling schemes. Each scheme allocates the 8 orders to the 5 riders in a different way. The order scheduling system sequentially inputs these 20 candidate schemes into the order scheduling model. The order scheduling model outputs a score for each scheme; for example, scheme A scores 8.5, scheme B scores 7.2, scheme C scores 9.0, and so on. The order scheduling system ranks and compares these output scheme scores, identifying the candidate scheme with the highest score. For example, scheme C, with a score of 9.0, ranks first, and is thus determined as the target order scheduling scheme.

[0064] Finally, the order dispatch system executes specific order allocation operations according to the target order dispatch plan. Continuing with the example above, it needs to issue the corresponding order instructions to the corresponding rider terminals according to the assignment relationship defined in plan C. For example, it notifies rider A to pick up food from restaurant A and deliver it to community B, and notifies rider B to pick up food from restaurant C, thereby completing the implementation of this round of dispatch decisions.

[0065] In this way, through the above process, the trained order scheduling model can quickly evaluate a large number of candidate solutions in a real-time environment and select the best one to execute, realizing a complete closed loop from offline learning to online application. This enables scheduling decisions to dynamically adapt to complex environments such as sudden changes in supply and demand and severe weather, avoiding the rigid performance of traditional rule engines when facing unexpected scenarios, and significantly improving the accuracy, efficiency and intelligence of order scheduling.

[0066] In summary, the logical process of the technical solution in the embodiments of this application is summarized as follows: Figure 2 As shown, Multiple platform samples are acquired, which are also sample order scheduling schemes. Each platform sample corresponds to a candidate order-rider matching scheme at a historical moment. For example, in sample S1, order O1 is assigned to rider C1, order O2 is assigned to rider C2, and in sample S2, order O3 is assigned to rider C3, and so on. For each sample, value characterization is first performed by calculating its feature indicators from multiple dimensions to form a platform value vector. This vector includes values ​​such as average route affinity m1, average delivery time m2, average delivery distance m3, average on-time rate m4, and average rider evaluation m5, thus obtaining the vector representation corresponding to each sample, such as vector1, vector2, vector3, etc. Subsequently, the implicit value inference stage begins. Based on these vectors, the maximum likelihood estimation method is used to jointly learn the implicit values ​​v1, v2, v3, etc. of each platform, the nonlinear response functions of each indicator, and noise parameters. Through optimization, the model achieves a better fit to the observed indicators, and finally, the implicit value of each platform is obtained, which is also known as the implicit quality score. Next, preference labels are constructed based on the learned latent values. By comparing the latent values ​​of different game hands (e.g., v1 > v2, v1 > v3, v2 > v3), preference pairs are automatically generated and their relative merits are marked. Finally, these preference pairs are used to perform Bradley-Terry preference learning, with the training objective of minimizing the negative log-likelihood loss. A function R(S) capable of scoring any game hand is learned as the order scheduling model. This function can be directly used to evaluate the merits of new scheduling schemes, thus completing the training process of the entire order scheduling model.

[0067] Thus, the technical solution in this application constructs a multi-indicator latent value modeling framework, treating multi-dimensional service indicators as noisy observations of scheduling latent value. By leveraging nonlinear latent variable models and maximum likelihood estimation, it inversely derives the objective latent value of each sample order scheduling scheme from historical data, thereby overcoming the limitations of traditional methods that rely on manually set weights. This significantly improves the multi-indicator fusion effect and achieves objective and automated value inference. Simultaneously, by introducing noise parameters into the model to accurately reflect the reliability differences of each indicator, it greatly enhances robustness to observation noise, enabling value assessment to adapt to the fluctuating characteristics of different indicators. Based on obtaining the latent value, the technical solution in this application further avoids the parameter optimization dilemma caused by manual weighted summation by automatically generating preference labels, allowing the model to flexibly adapt to dynamically changing scenarios. Finally, it employs the Bradley-Terry preference learning framework, using relative preference signals rather than labels to learn a reward function that can explain preferences, accommodating complex trade-offs between multiple objectives and solving the drawbacks of traditional methods that rely on manually constructed pseudo-rewards for explicit regression.

[0068] The method provided in this application obtains multiple sample scheduling schemes and evaluates them using multi-dimensional feature indicators to obtain multiple indicator scores. Then, it uses a pre-constructed likelihood estimation function to deeply mine these indicator scores and obtain implicit quality scores that reflect the true comprehensive quality of the schemes. This achieves the automatic extraction of objective and denoised scheduling effect measurements from multi-dimensional indicators, avoiding the subjectivity and rigid boundaries of relying on manually set rule weights. Furthermore, it automatically generates and labels preference pairs based on the implicit quality scores and uses these preference pairs to train an order scheduling model. This enables the order scheduling model to learn the inherent superiority-inferiority relationships between scheduling schemes from the data, effectively avoiding rule conflicts and local optima problems. In this way, it can dynamically adapt to complex environments such as sudden changes in supply and demand and abnormal weather during actual scheduling, improving the accuracy and efficiency of order scheduling and demonstrating good intelligence.

[0069] Furthermore, as Figure 1 To specifically implement the method, this application provides an order scheduling device, such as... Figure 3 As shown, the device includes: a first evaluation module 301, a second evaluation module 302, a preference pair generation module 303, a training module 304, and a scheduling module 305.

[0070] The first evaluation module 301 is used to obtain multiple sample order scheduling schemes and evaluate each sample order scheduling scheme using multi-dimensional feature indicators to obtain multiple indicator scores. The second evaluation module 302 is used to estimate the true quality of each sample order scheduling scheme by combining a pre-constructed likelihood estimation function with multiple index scores corresponding to each sample order scheduling scheme, so as to obtain the implicit quality score corresponding to each sample order scheduling scheme. The preference pair generation module 303 is used to organize the multiple sample order scheduling schemes according to the implicit quality score corresponding to each sample order scheduling scheme, obtain multiple preference pairs, and mark each preference pair; Training module 304 is used to train the order scheduling model using the labeled multiple preference pairs; The scheduling module 305 is used to call the order scheduling model to generate a target order scheduling scheme and execute the target order scheduling scheme when performing order scheduling.

[0071] In a specific application scenario, the first evaluation module 301 is used to evaluate the sample order scheduling scheme in multiple dimensions for each sample order scheduling scheme to obtain the multiple indicator scores, wherein the multiple indicator scores include one or more of dispatch interference scores, wave scores, and delivery object scores.

[0072] In specific application scenarios, the first evaluation module 301 is used to determine each sample delivery object to be dispatched according to the sample order scheduling scheme, dispatch the specified orders to each sample delivery object according to the sample order scheduling scheme and the orders currently accepted by each sample delivery object, generate an updated delivery path for each sample delivery object, query the current delivery path of each sample delivery object, and calculate the route compatibility between the updated delivery path and the current delivery path of each sample delivery object to obtain the dispatch interference score; and / or, dispatch the specified orders to each sample delivery object according to the sample order scheduling scheme to obtain the order dispatch interference score for each sample delivery object. After the sample order scheduling plan is completed, a delivery task list is generated for each sample delivery object. Based on the delivery task list corresponding to each sample delivery object, a global efficiency index is calculated for each sample delivery object, and the average of the global efficiency indices of all sample delivery objects is calculated to obtain the wave score. And / or, the feedback data on the delivery task list corresponding to each sample delivery object after the sample order scheduling plan is completed is estimated to obtain the estimated feedback value corresponding to each sample delivery object, and the average of the estimated feedback values ​​of all sample delivery objects is calculated to obtain the delivery object score.

[0073] In specific application scenarios, the device further includes: A function construction module is used to determine multiple preset service indicators and, in conjunction with the multiple sample order scheduling schemes and the multiple service indicators, establish an observation model. In the observation model, each sample order scheduling scheme corresponds to an unobservable latent quality variable, and each service indicator is represented as the observation result of the corresponding latent quality variable mapped by a nonlinear response function and superimposed with Gaussian noise. The variance of the Gaussian noise is used to characterize the observation reliability of the corresponding service indicator. Based on the observation model, for the multiple sample order scheduling schemes and the multiple service indicators, a log-likelihood function is constructed as the likelihood estimation function, using the latent quality score of each sample order scheduling scheme, the nonlinear response function parameter of each service indicator, and the noise variance of each service indicator as variables. The likelihood estimation function is used to quantify the joint probability distribution of the actual values ​​of each service indicator of each sample order scheduling scheme observed under the values ​​of the variables.

[0074] In a specific application scenario, the second evaluation module 302 is used to optimize the implicit quality score of each sample order scheduling scheme, the nonlinear response function parameters of each service indicator, and the noise variance of each service indicator by maximizing the likelihood estimation function and combining multiple indicator scores corresponding to each sample order scheduling scheme, so that the likelihood estimation function fits the actual observed data to a preset convergence condition; and outputs the implicit quality score of each sample order scheduling scheme obtained after optimization and convergence as the implicit quality score corresponding to each sample order scheduling scheme.

[0075] In a specific application scenario, the preference pair generation module 303 is used to compare the implicit quality scores corresponding to every two sample order scheduling schemes, combine two sample order scheduling schemes with a relationship between their implicit quality scores into a preference pair, so as to obtain the multiple preference pairs; identify the relationship between the implicit quality scores corresponding to the two sample order scheduling schemes in each preference pair, and mark each preference pair.

[0076] In a specific application scenario, the preference pair generation module 303 is used to, for each preference pair, take one sample order scheduling scheme in the preference pair as a first sample order scheduling scheme, and take the other sample order scheduling scheme in the preference pair other than the first sample order scheduling scheme as a second sample order scheduling scheme, wherein the implicit quality score corresponding to the first sample order scheduling scheme is greater than the implicit quality score corresponding to the second sample order scheduling scheme; generate a label indicating that the first sample order scheduling scheme is better than the second sample order scheduling scheme, and use the label to mark the preference pair.

[0077] In a specific application scenario, the training module 304 is used to construct the order scheduling model to be trained; calculate the difference between the implicit quality scores corresponding to the two sample order scheduling schemes in each preference pair; input each labeled preference pair and its corresponding difference into a preset normalized exponential function to obtain the predicted probability corresponding to each preference pair, wherein the predicted probability is used to indicate the probability that the sample order scheduling scheme with a high implicit quality score in the corresponding preference pair is better than the sample order scheduling scheme with a low implicit quality score; based on the predicted probabilities corresponding to the multiple preference pairs, a negative log-likelihood loss function is constructed, which is used to measure the difference between the actual superiority / inferiority relationship and the predicted probability in each preference pair; with minimizing the negative log-likelihood loss function as the optimization objective, the parameters of the order scheduling model are iteratively updated so that the scheme score output by the order scheduling model converges, thereby obtaining the trained order scheduling model.

[0078] In a specific application scenario, the scheduling module 305 is used to acquire real-time scheduling data at the current moment when scheduling orders. The real-time scheduling data includes information on orders to be assigned, information on available delivery objects, and environmental status information. Referring to the real-time scheduling data, multiple candidate order scheduling schemes are constructed, and these schemes are input into the order scheduling model. The order scheduling model calculates and outputs a score for each candidate scheme. Based on the score of each candidate scheme output by the order scheduling model, the scores of the multiple candidate schemes are compared, and the candidate scheme with the highest score is selected as the target order scheduling scheme. Order allocation is performed according to the target order scheduling scheme, assigning the corresponding orders to the corresponding delivery objects.

[0079] The apparatus provided in this application acquires multiple sample scheduling schemes and evaluates them using multi-dimensional feature indicators to obtain multiple indicator scores. Then, it uses a pre-constructed likelihood estimation function to perform deep mining on these indicator scores to obtain implicit quality scores that reflect the true comprehensive quality of the schemes. This achieves the automatic extraction of objective and denoised scheduling effect measurements from multi-dimensional indicators, avoiding the subjectivity and rigid boundaries of relying on manually set rule weights. Furthermore, it automatically generates and labels preference pairs based on the implicit quality scores and uses these preference pairs to train an order scheduling model. This enables the order scheduling model to learn the inherent superiority-inferiority relationships between scheduling schemes from the data, effectively avoiding rule conflicts and local optima problems. As a result, it can dynamically adapt to complex environments such as sudden changes in supply and demand and abnormal weather during actual scheduling, improving the accuracy and efficiency of order scheduling and demonstrating good intelligence.

[0080] It should be noted that other corresponding descriptions of the functional units involved in the order scheduling device provided in this application embodiment can be found by referring to... Figure 1 and Figure 2 The corresponding description in [the document] will not be repeated here.

[0081] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties.

[0082] The above embodiments and the technical features in the embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0083] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

[0084] In an exemplary embodiment, see Figure 4 The invention also provides a computer device including a bus, a processor, a memory, and a communication interface. It may also include an input / output interface and a display device, wherein the various functional units can communicate with each other via the bus. The memory stores a computer program, and the processor executes the program stored in the memory to perform the order scheduling method described in the above embodiments.

[0085] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the order scheduling method.

[0086] Through the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented in hardware or by using software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solution of this application can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, external hard drive, etc.) and includes several instructions to cause a computer device (such as a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0087] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of a preferred embodiment, and the modules or processes shown in the drawings are not necessarily essential for implementing this application.

[0088] Those skilled in the art will understand that the modules in the apparatus of the implementation scenario can be distributed within the apparatus of the implementation scenario as described, or they can be located in one or more apparatuses different from this implementation scenario, with corresponding changes. The modules of the above-described implementation scenario can be combined into one module, or they can be further divided into multiple sub-modules.

[0089] The serial numbers in this application are for descriptive purposes only and do not represent the superiority or inferiority of the implementation scenario.

[0090] The above disclosures are only a few specific implementation scenarios of this application. However, this application is not limited to these. Any variations that can be conceived by those skilled in the art should fall within the protection scope of this application.

Claims

1. An order dispatching method characterized by, include: Multiple sample order scheduling schemes are obtained, and each sample order scheduling scheme is evaluated by multi-dimensional feature indicators to obtain multiple indicator scores; By using a pre-constructed likelihood estimation function and combining multiple index scores corresponding to each sample order scheduling scheme, the true quality of each sample order scheduling scheme is estimated to obtain the implicit quality score corresponding to each sample order scheduling scheme. Based on the implicit quality score corresponding to each of the sample order scheduling schemes, the multiple sample order scheduling schemes are sorted out to obtain multiple preference pairs, and each preference pair is labeled. The order scheduling model is trained using the labeled multiple preference pairs, and when scheduling orders, the order scheduling model is invoked to generate a target order scheduling scheme and execute the target order scheduling scheme.

2. The method according to claim 1, characterized in that, The process of evaluating each sample order scheduling scheme using multi-dimensional feature indicators yields multiple indicator scores, including: For each of the sample order scheduling schemes, the sample order scheduling schemes are evaluated by feature indicators in multiple dimensions to obtain multiple indicator scores, wherein the multiple indicator scores include one or more of dispatch interference scores, wave scores, and delivery object scores.

3. The method according to claim 2, characterized in that, The evaluation of the sample order scheduling scheme based on feature indicators across multiple dimensions yields scores for these multiple indicators, including: The sample order scheduling scheme determines each sample delivery object to be dispatched, and according to the specified orders dispatched to each sample delivery object and the currently accepted orders of each sample delivery object, an updated delivery route is generated for each sample delivery object. The current delivery route of each sample delivery object is then queried, and the route compatibility between the updated delivery route and the current delivery route is calculated to obtain the dispatch interference score; and / or, According to the sample order scheduling scheme, orders are scheduled to the designated orders of each sample delivery object. A delivery task list is generated for each sample delivery object after the sample order scheduling scheme is completed. Based on the delivery task list for each sample delivery object, a global efficiency index is calculated for each sample delivery object, and the average global efficiency index of all sample delivery objects is calculated to obtain the wave score; and / or, The feedback data on the corresponding delivery task list for each sample delivery object after the sample order scheduling plan is completed is estimated to obtain the estimated feedback value for each sample delivery object, and the average value of the estimated feedback values ​​of all sample delivery objects is calculated to obtain the delivery object score.

4. The method according to claim 1, characterized in that, Before estimating the true quality of each sample order scheduling scheme by combining a pre-constructed likelihood estimation function with multiple indicator scores corresponding to each sample order scheduling scheme to obtain the implicit quality score corresponding to each sample order scheduling scheme, the method further includes: Multiple preset service indicators are determined, and an observation model is established by combining the multiple sample order scheduling schemes and the multiple service indicators. In the observation model, each sample order scheduling scheme corresponds to an unobservable latent quality variable, and each service indicator is represented as the observation result of the corresponding latent quality variable after being mapped by a nonlinear response function and superimposed with Gaussian noise. The variance of the Gaussian noise is used to characterize the observation reliability of the corresponding service indicator. Based on the observation model, for the multiple sample order scheduling schemes and the multiple service indicators, a log-likelihood function is constructed as the likelihood estimation function, using the implicit quality score of each sample order scheduling scheme, the nonlinear response function parameter of each service indicator, and the noise variance of each service indicator as variables. The likelihood estimation function is used to quantify the joint probability distribution of the actual values ​​of each service indicator of each sample order scheduling scheme observed under the values ​​of the variables.

5. The method according to claim 4, characterized in that, The process involves estimating the true quality of each sample order scheduling scheme by combining a pre-constructed likelihood estimation function with multiple indicator scores corresponding to each scheme, thereby obtaining an implicit quality score for each scheme. This includes: By maximizing the likelihood estimation function and combining it with multiple index scores corresponding to each sample order scheduling scheme, the implicit quality score of each sample order scheduling scheme, the nonlinear response function parameters of each service index, and the noise variance of each service index are solved and optimized, so that the likelihood estimation function fits the actual observation data to the preset convergence condition. The implicit quality score of each sample order scheduling scheme obtained after optimization convergence is used as the implicit quality score corresponding to each sample order scheduling scheme and output.

6. The method according to claim 1, characterized in that, The process involves organizing the multiple sample order scheduling schemes according to the implicit quality score corresponding to each sample order scheduling scheme, obtaining multiple preference pairs, and labeling each preference pair, including: The implicit quality scores corresponding to each pair of sample order scheduling schemes are compared, and the two sample order scheduling schemes with a relationship between their implicit quality scores are combined into a preference pair to obtain the multiple preference pairs. Identify the magnitude relationship between the implicit quality scores corresponding to the two sample order scheduling schemes in each preference pair, and label each preference pair.

7. The method according to claim 6, characterized in that, The process of identifying the implicit quality scores corresponding to the two sample order scheduling schemes in each preference pair and labeling each preference pair includes: For each preference pair, one sample order scheduling scheme in the preference pair is taken as the first sample order scheduling scheme, and the other sample order scheduling scheme in the preference pair other than the first sample order scheduling scheme is taken as the second sample order scheduling scheme, wherein the implicit quality score corresponding to the first sample order scheduling scheme is greater than the implicit quality score corresponding to the second sample order scheduling scheme. Generate a label indicating that the first sample order scheduling scheme is better than the second sample order scheduling scheme, and use the label to mark the preference pair.

8. The method according to claim 1, characterized in that, The process of training the order scheduling model using the labeled multiple preference pairs includes: Construct the order scheduling model to be trained; Calculate the difference between the implicit quality scores of the two sample order scheduling schemes in each preference pair, and input each labeled preference pair and its corresponding difference into a preset normalized exponential function to obtain the predicted probability of each preference pair. The predicted probability is used to indicate the probability that the sample order scheduling scheme with a high implicit quality score in the corresponding preference pair is better than the sample order scheduling scheme with a low implicit quality score. Based on the predicted probabilities corresponding to the multiple preference pairs, a negative log-likelihood loss function is constructed. The negative log-likelihood loss function is used to measure the difference between the actual superiority / inferiority relationship and the predicted probability in each preference pair. With minimizing the negative log-likelihood loss function as the optimization objective, the parameters of the order scheduling model are iteratively updated until the scheme score output by the order scheduling model converges, thereby obtaining the trained order scheduling model.

9. The method according to claim 1, characterized in that, The step of calling the order scheduling model to generate a target order scheduling plan and executing the target order scheduling plan during order scheduling includes: When scheduling orders, real-time scheduling data is obtained at the current moment. The real-time scheduling data includes information on orders to be assigned, information on available delivery objects, and information on environmental status. Based on the real-time scheduling data, multiple candidate order scheduling schemes are constructed, and the multiple candidate order scheduling schemes are input into the order scheduling model so that the order scheduling model calculates and outputs a scheme score for each candidate order scheduling scheme. Based on the scheme score of each candidate order scheduling scheme output by the order scheduling model, the scores of the multiple candidate order scheduling schemes are compared, and the candidate order scheduling scheme with the highest scheme score is selected as the target order scheduling scheme. The order allocation operation is performed according to the target order scheduling scheme, and the corresponding orders are assigned to the corresponding delivery objects.

10. An order scheduling device, characterized in that, include: The first evaluation module is used to obtain multiple sample order scheduling schemes and evaluate each sample order scheduling scheme using multi-dimensional feature indicators to obtain multiple indicator scores. The second evaluation module is used to estimate the true quality of each sample order scheduling scheme by combining a pre-constructed likelihood estimation function with multiple index scores corresponding to each sample order scheduling scheme, so as to obtain the implicit quality score corresponding to each sample order scheduling scheme. The preference pair generation module is used to organize the multiple sample order scheduling schemes according to the implicit quality score corresponding to each sample order scheduling scheme, obtain multiple preference pairs, and mark each preference pair; The training module is used to train the order scheduling model using the labeled multiple preference pairs; The scheduling module is used to call the order scheduling model to generate a target order scheduling scheme and execute the target order scheduling scheme when performing order scheduling.

11. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 9.

12. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 9.

Images