An artificial intelligence-based gig worker task scheduling method and system

Abstract

This application provides an AI-based method and system for scheduling gig workers' tasks, relating to the field of gig worker scheduling. It addresses the technical problems of existing technologies failing to effectively handle unexpected situations during task execution, leading to static allocation scheme failures, decreased service reliability, and low resource utilization. The method includes: acquiring information on tasks to be scheduled and gig worker information, constructing a composite scheduling utility function, and evaluating the comprehensive scheduling benefits of allocating tasks to gig workers; constructing a task-person bipartite graph using the comprehensive scheduling benefits as matching weights, and obtaining an initial task allocation scheme under the conditions of satisfying hard feasibility constraints and matching structure constraints; initiating task execution based on the initial task allocation scheme, and if the environmental state is abnormal during execution, initiating a dynamic re-matching mechanism to generate an updated task allocation scheme. This application is used in the gig worker task scheduling process.

Description

Technical Field

[0001] This application relates to the field of gig worker scheduling, and in particular to a gig worker task scheduling method and system based on artificial intelligence. Background Technology

[0002] With the rapid development of the sharing economy and platform-based gig economy, scenarios such as food delivery, instant retail, and on-demand services are placing higher demands on the real-time performance, robustness, and intelligence of task scheduling systems. Current technologies for gig worker task scheduling often employ static matching strategies, such as greedy algorithms, rule engines, or one-time maximum weight matching in a bipartite graph. These methods typically complete a one-time allocation at the beginning of the scheduling cycle based on task location, worker's current location, and simple service capability labels, lacking the ability to perceive and respond to dynamic changes in the environment during subsequent execution.

[0003] However, in real-world gig work scenarios, tasks are highly dynamic, subject to strong temporal and spatial constraints, and require multi-dimensional capabilities. Meanwhile, gig workers are highly mobile, and their status (such as fatigue, location, and availability) changes in real time. In the event of unforeseen circumstances, such as temporary worker departures, sudden traffic congestion, user changes to task addresses, or tasks facing timeout risks, static allocation schemes are highly susceptible to failure, leading to order fulfillment failures, resource mismatches, or a severe decline in user experience. Existing technologies lack effective online dynamic adjustment mechanisms, making it impossible to quickly generate feasible reallocation schemes while ensuring overall efficiency, thus making it difficult to balance the efficiency, reliability, and service resilience of the scheduling system.

[0004] Therefore, how to build an intelligent scheduling mechanism that can sense environmental disturbances in real time and adaptively generate optimized re-matching strategies while ensuring high-quality initial matching has become a core technical problem that urgently needs to be solved. Summary of the Invention

[0005] This application provides a task scheduling method and system for gig workers based on artificial intelligence, which solves the technical problems of existing technologies being unable to effectively cope with emergencies during task execution, resulting in the failure of static allocation schemes, decreased service reliability, and low resource utilization.

[0006] To achieve the above objectives, this application adopts the following technical solution: Firstly, an artificial intelligence-based task scheduling method for gig workers is provided, including: Obtain information on tasks to be scheduled and information on gig workers, and construct a composite scheduling utility function; evaluate the overall scheduling benefits of allocating tasks to gig workers based on the composite scheduling utility function; Using the comprehensive scheduling benefits as matching weights, a task-person bipartite graph is constructed. Under the conditions of satisfying hard feasibility constraints and matching structure constraints, the maximum weight matching problem of the bipartite graph is solved to obtain the initial task allocation scheme. Among them, the hard feasibility constraints include task time window constraints and personnel reachability constraints, and the matching structure constraints include that each task can be assigned to at most one gig worker and each gig worker can undertake at most one task. Task execution is initiated based on the initial task allocation scheme, and the environmental status is monitored in real time during execution. If the environmental status is abnormal, a dynamic re-matching mechanism is initiated with the current set of unfinished tasks and the status of temporary workers as input to generate an updated task allocation scheme. Among them, abnormal environmental status includes any assigned task that cannot be completed as originally planned due to personnel withdrawal, traffic abnormalities, task changes, or timeout risks.

[0007] Based on the above technical solutions, this application provides an AI-based task scheduling method for gig workers. In the gig economy scenario, tasks are highly dynamic, subject to strong spatiotemporal constraints, and require multi-dimensional capabilities. Meanwhile, personnel mobility is high and their states change in real time. Traditional static scheduling methods struggle to cope with sudden disturbances (such as personnel leaving, traffic congestion, and task changes), easily leading to order failures, resource waste, or a decline in user experience. Therefore, there is an urgent need for an intelligent scheduling mechanism that can achieve high-quality initial matching and possess online adaptive adjustment capabilities. This application constructs a composite scheduling utility function that integrates multi-dimensional factors such as space, skills, reputation, and fatigue to accurately quantify the quality of task-person matching. Based on this, it uses maximum-weighted bipartite graph matching to solve the initial allocation, ensuring global optimality and feasibility. More importantly, it introduces a dynamic re-matching mechanism based on reinforcement learning, which can quickly generate new solutions that balance the current state and long-term benefits when environmental anomalies are detected. This solution not only addresses the pain point of "unreliable one-time allocation," but also achieves a balance between scheduling efficiency, robustness, and service reliability through a two-stage architecture of "static optimization + dynamic repair." It is significantly superior to existing methods that rely solely on rules or pure static optimization, providing a feasible intelligent decision-making foundation for large-scale, high-concurrency gig platforms.

[0008] In conjunction with the first aspect above, in one possible implementation, the method for constructing the composite scheduling utility function includes: The task information to be scheduled includes the task's geographical location. Skills demand vector Time window and urgency weight ;in, Let i be the i-th task to be scheduled, where i is the index number of the task to be scheduled. Tasks to be scheduled Early start time, Tasks to be scheduled Late closing time; The personnel information includes the real-time location of casual workers. Capability profile vector Current fatigue index and historical performance rating ;in, Let j be the j-th gig worker, where j is the index number of the gig worker; The composite scheduling utility function Defined as: ; in, These are preset parameters for the spatial attenuation scale. This is a skill-dimensional covariance matrix estimated based on historical successful matching data, used to characterize the correlation and tolerance resilience among different skills. The total number of skill dimensions. It is the inverse of the covariance matrix of the skill dimension. To balance the weighting of skill matching and historical reputation, This is a parameter for adjusting fatigue sensitivity.

[0009] In conjunction with the first aspect above, in one possible implementation, the capability profile... The methods for obtaining it include: Extracting gig workers by analyzing platform user registration information and historical task completion records. Behavioral evidence across multiple preset skill dimensions, including professional qualification certificates, frequency of task type completion, user evaluation keywords, and platform certification level; Constructing a task-skill mapping matrix ;in, The total number of task categories. The total number of skill dimensions. Indicates the first Class of tasks requires the first Item skill; For each skill dimension Calculate the number of gig workers In the Overall ability score in each skill: ; in, Personnel Whether to hold the execution of the Minimum qualification certificate required for this skill; For personnel Complete the first within the historically preset timeframe. Total number of tasks of this type; This serves as an index when iterating through or taking extreme values ​​of all gig workers. To extract information about the first point from user comments based on natural language processing Emotional score for each skill; Preset weighting coefficients; This is the Sigmoid normalization function.

[0010] In conjunction with the first aspect above, in one possible implementation, the fatigue index... The methods for obtaining it include: Real-time data collection of gig workers Terminal sensor data and task execution logs, including the continuous working hours for the day. Number of tasks completed per unit of time Movement speed volatility and physiological indicators of heart rate variability; The sensor data and task execution logs are input into a pre-trained fatigue perception neural network model, which outputs the current fatigue index. The fatigue perception neural network model is a three-layer fully connected network, trained on a labeled dataset. The labels are the fatigue levels assessed by experts based on task timeout rate, frequency of operational errors, and overall evaluation.

[0011] In conjunction with the first aspect above, in one possible implementation, the historical performance score... The methods for obtaining it include: Based on gig workers Historical task performance records, calculating historical performance scores. : ; in, The number of historical tasks considered in the current time-sharing calculation of contract fulfillment. This serves as a sequence number index for historical tasks. As an indicator variable for whether it is completed on time, Rate the task for the user. For time decay weight, For the first The time of completion of this task. , To adjust the parameters.

[0012] In conjunction with the first aspect above, in one possible implementation, the process of constructing the task-person bipartite graph includes: Based on the aforementioned comprehensive scheduling benefit value As edge weights, construct a task-person bipartite graph. The vertex set consists of the set of tasks to be scheduled. Gathering with casual workers constitute, The total number of tasks to be scheduled. The total number of gig workers, side collection This includes all task-person pairings that meet the hard feasibility constraints, and each edge Weighted ; The hard feasibility constraints include: Personnel accessibility constraints: gig workers Capability Profile Vector In the mission The projection on the required skill dimension is not lower than the preset threshold; Task time window constraint: Travel time calculated based on a real-time traffic flow prediction model satisfy: ; Where t is the current time, The average working speed, The minimum duration required for the task to be executed.

[0013] In conjunction with the first aspect above, in one possible implementation, the method for obtaining the initial matching scheme includes: Based on the bipartite graph Construct the cost matrix For each feasible edge Its corresponding element is defined as ; A variant of the Hungarian algorithm with dynamic pruning is used to solve the minimum-cost maximum-weight matching problem on this bipartite graph, obtaining the result that maximizes total utility. Maximize the initial task allocation scheme In each iteration of the algorithm, before performing augmenting path search, the candidate edge set is dynamically pruned. The candidate edge set is the set of all task-person pairings that satisfy the hard feasibility constraints and matching structure constraints at the current scheduling time.

[0014] In conjunction with the first aspect above, in one possible implementation, the dynamic rematching mechanism includes the following matching process: Constructing the state space This is used to characterize the dynamic configuration of the current scheduling environment; where each state... Encodes the set of incomplete tasks Aggregator with available gig workers Joint information; Based on the state space Constructing action space ; among them, each action This corresponds to making a new set of task-person pairing decisions between the currently unassigned or reassigned subset of tasks and the available personnel subset; Define reward function Its value is in the state Next action The resulting increase in global system utility is the sum of the comprehensive scheduling benefits corresponding to all assigned tasks in the updated task allocation scheme. The state The vectorized representation is then input into a deep Q-network model, which outputs the action space. The Q value corresponds to each action, where Q is the cumulative reward expected to be obtained by taking each rematching action in the current state. Each time dynamic rematch is triggered, the action with the largest Q value is selected as the optimal rematch strategy based on the Q value output by the deep Q-network model, and an updated task allocation scheme is generated based on this action. .

[0015] In conjunction with the first aspect above, in one possible implementation, the method for iteratively optimizing the Q-value of the deep Q-network model includes: ; in, This is the current state. To perform the action The new state afterwards; For the new state Candidate actions to be executed, Main network parameters, For target network parameters; For learning rate, Discount factor; Indicates the state Take action below The expected cumulative reward.

[0016] Secondly, this application provides an artificial intelligence-based task scheduling system for gig workers, comprising: an analysis module, an allocation module, and an execution module; wherein, the analysis module is used to acquire information on tasks to be scheduled and information on gig workers, and construct a composite scheduling utility function; based on the composite scheduling utility function, it evaluates the comprehensive scheduling benefits of allocating tasks to gig workers; the allocation module is used to construct a task-person bipartite graph using the comprehensive scheduling benefits as matching weights, and under the conditions of satisfying hard feasibility constraints and matching structure constraints, solves the maximum weight matching problem of the bipartite graph to obtain an initial task allocation scheme; the execution module is used to start task execution based on the initial task allocation scheme, and monitors the environmental status in real time during execution. If the environmental status is abnormal, it uses the current set of unfinished tasks and the status of gig workers as inputs to start a dynamic re-matching mechanism to generate an updated task allocation scheme.

[0017] This application provides an AI-based task scheduling method and system for gig workers, achieving a unified approach of high-precision initial matching and real-time dynamic re-optimization, significantly improving the scheduling efficiency and service reliability of gig platforms in complex and highly dynamic environments. The method integrates multi-dimensional heterogeneous information such as task geographic location, skill requirements, time windows, urgency, and personnel capability profiles, historical performance scores, and fatigue status to construct a composite scheduling utility function, accurately quantifying the comprehensive matching benefit of each task-person pair. Then, using this benefit as a weight, and under the premise of strictly satisfying hard feasibility constraints such as task time windows and personnel accessibility, as well as a "one-to-one" matching structure, a maximum-weighted bipartite graph matching algorithm is used to solve for the globally optimal initial allocation scheme. More importantly, the system continuously monitors the operational status during task execution. Once environmental disturbances such as personnel exit, traffic anomalies, task changes, or timeout risks are detected, a dynamic re-matching mechanism is immediately triggered—utilizing a deep reinforcement learning model to quickly generate an updated allocation scheme that balances immediate utility and long-term system performance based on the current set of incomplete tasks and the real-time status of available personnel. Compared to traditional static scheduling or rule-driven methods, this solution not only avoids service interruptions caused by sudden anomalies, but also continuously optimizes decision quality through data-driven adaptive learning capabilities, effectively balancing scheduling efficiency, resource utilization, and worker experience, providing core technical support for large-scale gig platforms that combines intelligence, robustness, and scalability.

[0018] It should be understood that the descriptions of technical features, technical solutions, beneficial effects, or similar language in this application do not imply that all features and advantages can be achieved in any single embodiment. Rather, it is understood that the description of a feature or beneficial effect means that a specific technical feature, technical solution, or beneficial effect is included in at least one embodiment. Therefore, the descriptions of technical features, technical solutions, or beneficial effects in this specification do not necessarily refer to the same embodiment. Furthermore, the technical features, technical solutions, and beneficial effects described in this embodiment can be combined in any suitable manner. Those skilled in the art will understand that embodiments can be implemented without one or more specific technical features, technical solutions, or beneficial effects of a particular embodiment. In other embodiments, additional technical features and beneficial effects may be identified in specific embodiments that do not embody all embodiments. Attached Figure Description

[0019] Figure 1 A system architecture diagram of an artificial intelligence-based task scheduling system for gig workers provided for embodiments of this application; Figure 2 A flowchart illustrating an artificial intelligence-based task scheduling method for gig workers, provided as an embodiment of this application; Figure 3 This is a flowchart illustrating a matching method for a dynamic matching mechanism provided in an embodiment of this application. Detailed Implementation

[0020] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] The task scheduling method for gig workers based on artificial intelligence provided in this application embodiment can be applied to, for example... Figure 1 The system shown is an AI-based task scheduling system for gig workers. The system includes an analysis module, an allocation module, and an execution module. The analysis module is used to obtain information on tasks to be scheduled and information on gig workers, and to construct a composite scheduling utility function; based on the composite scheduling utility function, the comprehensive scheduling benefits of allocating tasks to gig workers are evaluated. The allocation module is used to construct a task-person bipartite graph with the comprehensive scheduling benefits as the matching weight, and solve the maximum weight matching problem of the bipartite graph under the conditions of satisfying hard feasibility constraints and matching structure constraints to obtain the initial task allocation scheme. The execution module is used to start task execution based on the initial task allocation scheme and monitor the environment status in real time during execution. If the environment status is abnormal, the dynamic re-matching mechanism is started with the current set of unfinished tasks and the status of gig workers as input to generate an updated task allocation scheme.

[0022] To address the technical problems of existing technologies failing to effectively handle unforeseen circumstances during task execution, leading to the failure of static allocation schemes, decreased service reliability, and low resource utilization, this application provides an artificial intelligence-based task scheduling method for gig workers. The method includes: Obtain information on tasks to be scheduled and information on gig workers, and construct a composite scheduling utility function; evaluate the overall scheduling benefits of allocating tasks to gig workers based on the composite scheduling utility function; Using the comprehensive scheduling benefit as the matching weight, a task-person bipartite graph is constructed. Under the conditions of satisfying the hard feasibility constraints and the matching structure constraints, the maximum weight matching problem of the bipartite graph is solved to obtain the initial task allocation scheme. The hard feasibility constraints include task time window constraints and personnel reachability constraints. The matching structure constraints include that each task can be assigned to at most one gig worker and each gig worker can undertake at most one task. Task execution is initiated based on the initial task allocation scheme, and the environmental status is monitored in real time during execution. If the environmental status is abnormal, a dynamic re-matching mechanism is initiated with the current set of unfinished tasks and the status of temporary workers as input to generate an updated task allocation scheme. Among them, abnormal environmental status includes any assigned task that cannot be completed as originally planned due to personnel withdrawal, traffic abnormalities, task changes, or timeout risks.

[0023] Based on this, this solution significantly improves the efficiency and robustness of gig worker scheduling through a two-stage mechanism of "static optimization + dynamic repair." A composite scheduling utility function is constructed, integrating task attributes (location, time window, skill requirements) and personnel status (ability profile, reputation, fatigue level) to accurately quantify matching benefits. Using this function as weights, and under hard constraints such as time windows and reachability, as well as "one-to-one" structural constraints, the maximum weighted bipartite graph matching is solved to obtain a high-quality initial allocation scheme. More importantly, the environmental status is monitored in real time during task execution. Once disturbances such as personnel exit, traffic anomalies, task changes, or timeout risks occur, a dynamic re-matching mechanism is immediately triggered. Based on the real-time status of currently unfinished tasks and available personnel, a feasible new scheme is quickly generated. This method effectively overcomes the shortcomings of traditional static scheduling in handling dynamic disturbances, balancing global optimization and local adaptation, and significantly improving service reliability, resource utilization, and system resilience.

[0024] like Figure 2 As shown in the figure, an embodiment of this application provides a task scheduling method for gig workers based on artificial intelligence, including: S201. Obtain the set of tasks to be scheduled. Gathering with casual workers , The total number of tasks to be scheduled. This includes the total number of gig workers, as well as information on tasks to be scheduled and information on gig workers.

[0025] The information on tasks to be scheduled includes the task's geographical location. Skills demand vector Time window and urgency weight Where i is the index number of the task to be scheduled. For the i-th task to be scheduled, Tasks to be scheduled Earliest start time Tasks to be scheduled Latest end time.

[0026] It should be noted that geographical location and time window together define the accessibility of the task in the spatiotemporal dimension, and need to be dynamically verified in conjunction with real-time traffic data; the skill demand vector expresses the multidimensional requirements of the task on personnel capabilities in a continuous and quantifiable way, and is the key basis for achieving accurate matching with the capability profile of gig workers; the urgency weight reflects the task priority and is used to guide the resource allocation of the utility function.

[0027] S202. Construct a composite scheduling utility function; evaluate the overall scheduling benefits of allocating tasks to gig workers based on the composite scheduling utility function.

[0028] S203. Using the comprehensive scheduling benefit as the matching weight, construct a task-personnel bipartite graph, and under the conditions of satisfying the hard feasibility constraint and the matching structure constraint, solve the maximum weight matching problem of the bipartite graph to obtain the initial task allocation scheme.

[0029] Among them, the hard feasibility constraints include task time window constraints and personnel accessibility constraints, and the matching structure constraints include that each task can be assigned to at most one gig worker and each gig worker can undertake at most one task.

[0030] S204. Start task execution based on the initial task allocation scheme, and monitor the environment status in real time during execution. If the environment status is abnormal, start the dynamic re-matching mechanism with the current set of unfinished tasks and the status of temporary workers as input to generate an updated task allocation scheme.

[0031] Among them, abnormal environmental conditions include any assigned task that cannot be completed as originally planned due to personnel withdrawal, traffic abnormalities, task changes, or risk of exceeding time limits.

[0032] Based on the above technical solutions, this application provides an AI-based task scheduling method for gig workers. In the gig economy environment, task demands change rapidly, personnel status is highly dynamic, and tasks often have strict time windows, geographical locations, and skill requirements. Traditional scheduling methods typically perform only one-time static allocation, which cannot cope with frequent anomalies during execution (such as riders temporarily quitting, sudden traffic congestion, users changing addresses, or tasks facing timeout risks), easily leading to service failures, resource mismatches, or deteriorated user experience. Therefore, there is an urgent need for an intelligent scheduling mechanism that combines high-quality initial matching with real-time adaptive adjustment capabilities. This solution constructs a composite scheduling utility function that comprehensively considers spatial accessibility, skill matching degree, historical reputation, and personnel fatigue status to scientifically evaluate the matching value of each task-person pair. Using this as weight, under the premise of satisfying hard constraints such as time windows and capability accessibility, and a "one-to-one" matching structure, it solves the maximum weighted bipartite graph matching to obtain a globally optimal initial allocation scheme. More importantly, the system continuously monitors the environment during task execution. Once an anomaly is detected, a dynamic re-matching mechanism is immediately triggered. Based on the real-time status of currently incomplete tasks and available personnel, an updated feasible allocation plan is quickly generated. This two-stage architecture of "static optimization + dynamic repair" effectively balances scheduling efficiency, stability, and robustness, significantly improving the service capabilities and system resilience of the gig platform in complex real-world scenarios.

[0033] In one possible implementation of this application embodiment, the above-mentioned S202 can be specifically described as follows: Constructing a composite scheduling utility function : ; in, For spatial coupling terms, These are preset parameters for the spatial attenuation scale. For the ability-reputation item, This is a skill-dimensional covariance matrix estimated based on historical successful matching data, used to characterize the correlation and tolerance resilience among different skills. The total number of skill dimensions. It is the inverse of the covariance matrix of the skill dimension. For state adjustment items, To balance the weighting of skill matching and historical reputation, This is a parameter for adjusting fatigue sensitivity.

[0034] In some implementations, urgency weight The system is dynamically generated using a rule engine or machine learning model based on multi-dimensional features such as task type, user level, remaining time window, and whether it is a paid on-time delivery service, and is updated in real time as the task is executed; spatial decay scale. It can adaptively adjust based on geographical information such as urban traffic density and regional average driving speed. For example, it can use a smaller value in the city center to improve location accuracy and a larger value in the suburbs to enhance coverage; skill covariance matrix Continuously updated through an online learning mechanism, and using historical data from the most recent N successful matches for empirical estimation, ensuring it reflects the actual matching preferences of the current platform; fatigue index The calculation is based on a comprehensive assessment of factors such as continuous working hours, order density, and travel intensity, and a sliding window mechanism is introduced to prevent misjudgment due to short-term fluctuations.

[0035] In some implementations, the weighting coefficient This is used to dynamically balance the contribution ratio of skill matching degree and historical performance reputation to total utility. Its value can be configured according to task type or platform operation strategy: for example, in high-skill-barrier tasks (such as home appliance installation, medical care companionship), The value is set to a higher value (e.g., 0.7~0.9) to prioritize capability adaptability; however, in scenarios with high standardization and strong fault tolerance (such as ordinary food delivery), It is set to a lower value (e.g., 0.3~0.5) to place greater emphasis on the staff's historical service performance. Furthermore, It can also be based on an online learning mechanism to automatically adjust according to historical matching success rates and user satisfaction feedback, achieving data-driven adaptive optimization.

[0036] Correspondingly, fatigue sensitivity adjustment parameters This is used to control the inhibitory effect of personnel's current fatigue state on task allocation. Its value is related to the platform's security policy and labor compliance requirements: during peak hours or under severe weather conditions such as high temperature, rain, or snow, The value is dynamically increased (e.g., from the default value of 0.2 to 0.4~0.6) to significantly reduce the probability of highly fatigued personnel accepting orders; while during periods of sufficient capacity and lower task urgency, The efficiency can be appropriately reduced to maintain the overall system scheduling efficiency. In a preferred embodiment, Real-time fatigue index of personnel Linkage, when Automatic trigger when the preset threshold is exceeded The tiered amplification mechanism forces and restricts excessive continuous order taking, thereby ensuring service quality and staff health.

[0037] It should be noted that the composite scheduling utility function is not simply the sum of its factors, but rather employs a product structure to deeply integrate four dimensions: spatial accessibility, skill matching, historical reputation, and fatigue state. Specifically, the spatial term is modeled using a Gaussian kernel function to ensure rapid decay of long-distance matching; the skill term uses Mahalanobis distance, considering not only the magnitude of skill differences but also the correlation between skills (e.g., the strong correlation between "cycling" and "navigation") to improve matching robustness; and the reputation term... Positive incentives are introduced to encourage high-quality service, while a fatigue term acts as a determinant to avoid overworking personnel, thus achieving a balance between efficiency and sustainability. Therefore, this function is both mathematically rigorous and aligned with the actual operational needs of the gig economy.

[0038] For example, suppose a food delivery task Located in the city center, rider At a distance of 500 meters, the skill requirement and ability matching degree is 0.85, the historical performance score is 0.92, and the current fatigue index is 0.6. If set... , , , Then the spatial term Skills Fatigue inhibition term Ultimate utility This value can be used as input for subsequent matching algorithms, and the optimal allocation is determined after comparison with other pairings.

[0039] Based on the above technical solution, the constructed composite scheduling utility function comprehensively considers four key factors: spatial distance, skill matching, historical reputation, and fatigue status. This effectively solves the problems of low efficiency, poor service quality, and worker overwork caused by traditional task allocation that "relies solely on distance or simple matching." In the gig economy scenario, tasks are highly dynamic, subject to strong spatiotemporal constraints, and require multi-dimensional capabilities. Ignoring skill relevance, historical performance, and fatigue can easily lead to resource mismatch, increased user complaints, and health risks for workers. This solution uses Mahalanobis distance to characterize skill matching and employs a covariance matrix... Capture the correlation between different skills (such as the strong correlation between cycling and navigation) to improve matching robustness; introduce historical performance scores. As a reputation incentive, it promotes high-quality service; it also includes a fatigue suppression feature. Avoid overusing highly productive personnel. At the same time, use urgency weighting... This method achieves priority scheduling while balancing fairness and efficiency. It organically integrates multi-dimensional information into a unified mathematical framework, possessing advantages such as strong interpretability, adjustable parameters, and adaptability to complex scenarios, significantly improving the intelligence level and sustainable operation capability of the scheduling system.

[0040] Among them, ability profile The method for obtaining [the information] includes the following steps: S1. Extract gig workers by using platform user registration information and historical task completion records. Behavioral evidence across multiple preset skill dimensions, including professional qualification certificates, frequency of task completion, user review keywords, and platform certification level; S2. Constructing the Task-Skill Mapping Matrix ;in, The total number of task categories. The total number of skill dimensions. Indicates the first Class of tasks requires the first Item skill; For each skill dimension Calculate the number of gig workers In the Overall ability score in each skill: ; ability portrait The acquisition method includes two core steps: S1 is to extract raw evidence of individuals' skills across preset dimensions based on multi-source behavioral data; S2 is to obtain this evidence through a task-skill mapping matrix. The experience gained from completing tasks is aggregated into a unified skills space, and a comprehensive ability score is calculated by combining multi-dimensional information such as qualification certification and user evaluation. in, Personnel Whether to hold the execution of the Minimum qualification certificate required for this skill; For personnel Complete the first within the historically preset timeframe. Total number of tasks of this type; This serves as an index when iterating through or taking extreme values ​​of all gig workers. To extract information about the first point from user comments based on natural language processing Emotional score for each skill; Preset weighting coefficients are used to balance capability evidence from different sources; The Sigmoid normalization function, The composite score is compressed into the (0,1) interval to form a standardized ability score.

[0041] In some implementation methods, such as food delivery, preset skill dimensions are used. This includes, but is not limited to, "cycling ability," "navigation proficiency," "communication skills," "tool usage," and "area familiarity," with each skill dimension corresponding to a quantifiable behavioral indicator; task-skill mapping matrix. The platform operator statically configures the system according to business rules, for example, setting that "food delivery" tasks require two skills: "cycling" and "navigation." The time window for historical task completion records can be set to the most recent 90 days to ensure that the competency profile reflects the employee's recent performance; user review sentiment analysis uses a pre-trained language model (such as BERT) for fine-grained sentiment recognition, supporting keyword extraction and scoring of phrases such as "good service attitude," "arrived on time," and "poor attitude" in Chinese contexts; weighting coefficients It can be dynamically adjusted according to different task types. For example, the weight of qualification can be increased in high-risk tasks, and the weight of emotional score can be increased in service-oriented tasks.

[0042] It should be noted that the ability profile construction method does not simply use the number of historical tasks as a proxy variable for ability. Instead, it realizes the cross-task transfer and abstract expression of "task experience" to "general skills" through a task-skill mapping mechanism. That is, multiple completions of the same type of task are aggregated into experience accumulation of skills required for that type of task, thereby avoiding the fragmentation problem of ability assessment caused by the complexity of task types. At the same time, the method integrates three heterogeneous information sources: hard qualifications (certificates), soft experience (completion frequency), and subjective feedback (emotional score), forming a multimodal, interpretable, and robust comprehensive ability representation. In addition, the output is normalized by the Sigmoid function, which makes the ability score have good comparability and consistency, making it easy to use as input features in the matching algorithm.

[0043] For example, suppose a rider Holding a motorcycle driver's license (corresponding to "riding ability" skill), completed 80 food delivery tasks (classified as "delivery tasks") in the past 90 days, and received 50 user reviews, 45 of which mentioned "on-time delivery" and "good attitude." NLP analysis yielded a sentiment score. If set , , And in "delivery" tasks The overall score for their "cycling ability" is calculated as follows: ; Assuming the maximum number of completions is 100, then: ; This value represents the rider's standardized ability score in the "riding ability" dimension, which can be used for subsequent matching calculations with task requirements.

[0044] Based on the aforementioned technical solutions, this approach effectively addresses the core problems of traditional gig economy scheduling—namely, "one-sided capability assessment, low matching accuracy, and inability to generalize across tasks"—by constructing a multi-source capability profile that integrates qualifications, task experience, and user evaluations, and by introducing a task-skill mapping mechanism. In the gig economy context, with diverse task types and high personnel mobility, relying solely on simple tags or historical order volume can easily lead to skill mismatches, service failures, or resource waste. This solution employs a structured mapping and weighted fusion strategy to abstract discrete task experience into a general skill dimension, achieving uniformity and transferability in capability representation. Simultaneously, it combines real-time sentiment analysis and normalized scoring to enhance the dynamics and credibility of the profile. This method not only significantly improves task matching accuracy and user satisfaction but also provides the platform with an interpretable, configurable, and scalable capability assessment foundation, supporting the efficient operation of highly complex, large-scale real-time scheduling systems.

[0045] Fatigue Index The method for obtaining [the information] includes the following steps: Q1. Real-time data collection of gig workers Terminal sensor data and task execution logs, including the continuous working hours for the day. Number of tasks completed per unit of time Movement speed volatility And physiological indicators of heart rate variability, reflecting the regulatory capacity of the autonomic nervous system; Q2. Input the sensor data and task execution logs into the pre-trained fatigue perception neural network model, and output the current fatigue index. .

[0046] The fatigue perception neural network model is a three-layer fully connected network, trained on a labeled dataset. The labels are generated by experts based on task timeout rate, frequency of operational errors, and overall fatigue level assessment.

[0047] In some implementations, sensor data is collected through devices such as smart bracelets, mobile phone GPS, and accelerometers, with a sampling frequency set to 1-5 Hz per second to ensure high-precision capture of behavioral dynamics. The task execution log includes order acceptance time, departure time, arrival time, and route locations, which are automatically recorded by the platform system. The fatigue perception neural network model has a three-layer fully connected network structure, with an input layer dimension of 4 (corresponding to four features), 64 and 32 hidden layer nodes respectively, and an output layer using the Sigmoid activation function to ensure that the output is in the [0,1] interval. The model is trained using the Adam optimizer and cross-entropy loss function, and an early stopping mechanism is used to prevent overfitting. In a preferred implementation, the model supports online fine-tuning, continuously updating parameters using newly collected fatigue samples to improve long-term adaptability.

[0048] It should be noted that the fatigue index acquisition method described above does not rely on a single indicator (such as working hours) for simple threshold judgment. Instead, it integrates multimodal data on behavior, physiology, and task performance, and uses a deep learning model to achieve nonlinear mapping, thereby more accurately depicting the true fatigue state of personnel. This method avoids the subjectivity and lag of traditional "manual assessment" or "fixed rules," and has advantages such as strong real-time performance, high robustness, and good scalability. At the same time, by introducing expert annotations to build a high-quality training dataset, it ensures that the model output has clinical significance and business consistency, providing a reliable basis for subsequent scheduling decisions.

[0049] For example, suppose a rider Currently, I have been working continuously for 4 hours. On average, one order is completed every 10 minutes. The movement speed fluctuation rate was 0.35 (high, indicating frequent sudden braking / detours), and the heart rate variability was 40 ms (below normal, suggesting high stress). After inputting these features into the fatigue-sensing neural network, the model output... This indicates that the individual is in a state of moderate to high fatigue. Based on this, the dispatch system can lower their order priority or recommend rest to avoid safety incidents or negative service reviews caused by fatigue.

[0050] Based on the above technical solutions, this method achieves accurate and real-time assessment of fatigue indices for gig workers through multi-source fusion and neural network modeling. This effectively solves key problems in traditional scheduling systems that rely on simple rules (such as working time thresholds), leading to delayed fatigue identification, high misjudgment rates, and an inability to reflect the true state of the workforce. In high-intensity, high-mobility gig work scenarios, worker fatigue directly impacts service safety and quality. Failure to detect and intervene promptly can easily result in traffic accidents, order delays, or user complaints. This solution utilizes multi-dimensional features such as sensor data (e.g., heart rate variability), behavioral logs (task density, speed fluctuations), and working time, performing nonlinear fusion through a pre-trained neural network. This not only improves the accuracy and timeliness of fatigue assessment but also possesses adaptive and scalable capabilities. This method provides reliable health constraints for dynamic scheduling, balancing efficiency and worker rights, and serves as a core support for building a safe and sustainable intelligent scheduling system.

[0051] Historical performance rating : Based on gig workers Historical task performance records, calculating historical performance scores. It is based on gig workers The comprehensive credit index calculated from historical completion records; ; in, The number of historical tasks considered in the current time-sharing calculation of contract fulfillment. This serves as a sequence number index for historical tasks. As an indicator variable for whether it is completed on time, Rate the task for the user. The time decay weight reflects the greater impact of recent tasks on the current score. For the first The time of completion of this task. To adjust the parameters that correlate the importance of on-time completion with user ratings; The adjustment parameter is used to control the time decay rate.

[0052] In some implementations, the number of historical tasks The score can be set to the total number of tasks completed in the last 90 days, or dynamically updated using a sliding window mechanism, to ensure that the score reflects the staff's recent service capabilities; time decay weight. decay rate Configure according to business scenarios, such as boosting during peak hours. Values ​​that make recent performance more influential; parameters This can be adjusted according to different task types, such as increasing the weight of timely completion for time-sensitive tasks (e.g., food delivery). Increase the weight of user ratings for service experience-based tasks (such as housekeeping). In a preferred embodiment, the scoring module supports online updates, recalculating each time a new task is completed. This ensures its real-time and dynamic nature.

[0053] It should be noted that the historical performance scores mentioned above are not simply averaged, but rather a weighted average that integrates both timeliness of task completion and service quality, and incorporates a time decay mechanism. This ensures that recent performance has a greater impact on the current score, avoiding long-term misjudgments due to early low-quality service. This method retains the cumulative effect of historical experience while also being responsive to changes in behavior, effectively improving the accuracy and fairness of the scoring. Furthermore, adjustable parameters... and This enables flexible adaptation to different business scenarios, enhancing the system's versatility and controllability.

[0054] For example, suppose a rider Five tasks were completed in the past 30 days: the first was completed on time (score 0.8), the second was completed late (score 0.6), the third was completed on time (score 0.9), the fourth was completed on time (score 0.7), and the fifth was completed on time (score 0.9). Let... , The current time is The completion times for each task are as follows: .but: After calculating the weighted sum, we get This indicates that the individual's overall performance is good and they are suitable for priority assignment of high-value tasks.

[0055] Based on the above technical solution, by introducing a time decay weighting mechanism and integrating multi-dimensional performance indicators, the shortcomings of traditional credit scoring, such as "equal weighting of historical data, neglect of timeliness, and reliance on user scores," are effectively solved. In gig worker scheduling scenarios, personnel service capabilities exhibit significant time variability. If only a simple average is used, recent improvements may be masked by early negative records, or high-quality personnel may be misjudged due to occasional negative reviews, affecting matching fairness and system efficiency. This solution combines on-time completion rate with user scores in a weighted manner, giving higher weight to recent tasks, making the score more accurately reflect the current service level. Simultaneously, adjustable parameters support flexible adaptation to different business scenarios. This method not only improves the accuracy and timeliness of credit assessment but also enhances the incentive compatibility of the scheduling system, encouraging personnel to continuously provide high-quality services, thus forming a positive cycle and laying a reliable foundation for high-precision and robust intelligent matching.

[0056] In one possible implementation of this application embodiment, the above-mentioned S203 can be specifically implemented by the following S301, S302 and S303, which are described in detail below: S301, based on the comprehensive scheduling revenue value As edge weights, construct a task-person bipartite graph. .

[0057] The vertex set consists of the set of tasks to be scheduled. It is composed of groups of casual laborers; edge set This includes all task-person pairings that meet the hard feasibility constraints, and each edge Weighted ; Hard feasibility constraints include: Personnel accessibility constraints: gig workers Capability Profile Vector In the mission The projection on the required skill dimension is not lower than the preset threshold; Task time window constraint: Travel time calculated based on a real-time traffic flow prediction model satisfy: ; Where t is the current time, The average working speed, The minimum duration required for the task to be executed.

[0058] In some implementations, the capability profile projection threshold is dynamically set according to the task type. For example, high-skill tasks (such as home appliance repair) require a projection of no less than 0.8, while ordinary delivery tasks can be set to 0.6. The real-time traffic flow prediction model uses a deep learning-based spatiotemporal graph neural network (ST-GNN), taking historical road conditions, weather, events, and other data as input, and outputting the estimated travel time from the personnel's current location to the task location. Average operational movement speed... The system is configured according to task type and regional characteristics, such as 15km / h for urban cycling and 5km / h for walking. In a preferred embodiment, the bipartite graph construction process supports incremental updates. When a new task arrives or the status of personnel changes, only the affected edges are re-evaluated, thereby improving the system's response efficiency.

[0059] It should be noted that the bipartite graph construction method described above does not simply connect all tasks and personnel, but rather filters out truly "feasible" matching candidates through double hard constraints, ensuring that subsequent matching algorithms only search within the effective space and avoid ineffective computation. This method transforms the complex multidimensional scheduling problem into a standard graph theory structure, making it easy to solve efficiently using mature maximum weight matching algorithms (such as the Hungarian algorithm and Kuhn-Munkres). At the same time, by combining real-time traffic prediction and dynamic capability assessment, it achieves unified modeling of "physical reachability" and "capability adaptation," significantly improving the feasibility and implementability of the scheduling scheme.

[0060] For example, suppose a food delivery task Located in the city center, the latest closing time is 12:30, and the execution time is 10 minutes; riders The current location is 2 kilometers from the mission point. The traffic prediction model estimates the travel time to be 12 minutes, with an average speed of 15 km / h. If... minutes, then the total time is If the current time is 12:00, the completion time is 12:22, which is earlier than 12:30, satisfying the time window constraint. Simultaneously, its ability profile projected onto the "delivery" skill is 0.75, higher than the preset threshold of 0.6, also satisfying the skill constraint. Therefore, edge Added to edge set and assign weights , and participate in subsequent matching calculations.

[0061] Based on the above technical solution, by constructing a task-person bipartite graph that integrates hard feasibility constraints, the core problems of "blind matching, generation of infeasible solutions, and waste of computational resources" in traditional scheduling systems are effectively solved. In dynamic gig economy scenarios, tasks have strict spatiotemporal windows and skill requirements. If invalid pairs are not filtered in advance, the matching algorithm may output solutions that seem optimal but are actually unexecutable (such as personnel not arriving on time or lacking the necessary capabilities), leading to order failures or a decline in user experience. This solution introduces dual constraints of capability profile projection threshold and time window verification based on real-time traffic prediction during the graph construction phase, ensuring that the edge set only contains candidate pairs that are physically reachable and capability-matched. At the same time, the problem is transformed into a standard bipartite graph structure, which facilitates the invocation of efficient maximum weight matching algorithms. This method not only significantly improves the feasibility and reliability of scheduling results, but also reduces the search space and improves solution efficiency through pruning, providing a feasible and robust technical foundation for large-scale real-time intelligent scheduling.

[0062] S302, Based on bipartite graph Construct the cost matrix .

[0063] For each feasible edge Its corresponding element is defined as Cost matrix Based on the aforementioned task-personnel bipartite graph Constructed to transform the maximum weight matching problem into a minimum cost allocation problem; for each feasible edge Its corresponding element is defined as That is, the negative value of the composite scheduling utility function; the unconnected elements in this matrix (i.e., infeasible pairs) are usually set to Or a maximum value, to ensure that the algorithm does not select invalid matches.

[0064] In some implementations, the cost matrix supports sparse storage, storing only non-infinite elements to reduce memory usage and computational complexity. When the number of tasks or personnel changes dynamically, an incremental update mechanism can be used to modify only the affected rows or columns, avoiding the need to rebuild the entire matrix. In a preferred implementation, the matrix is ​​used as input to the Hungarian algorithm or the Kuhn-Munkres algorithm to solve the maximum weight bipartite graph matching problem. Furthermore, the matrix can be normalized before solving to improve numerical stability and convergence speed.

[0065] It should be noted that the above cost matrix construction method is not simply to invert the utility value, but rather to achieve a standard optimization problem mapping from "maximizing revenue" to "minimizing cost" through mathematical transformation, so that existing mature combinatorial optimization algorithms can be directly applied; this method maintains the relative size relationship of the original utility function to ensure that the optimal solution remains unchanged; at the same time, by setting infeasible edges to infinity, illegal matching is forcibly excluded, ensuring the feasibility and compliance of the final allocation scheme.

[0066] For example, suppose a certain task With rider The combined effect is The corresponding cost element is If the other pair If the time window constraint is not met, then After constructing the complete matrix, the Hungarian algorithm is called to solve for the minimum total cost allocation, and the result corresponds to the maximum total utility matching scheme.

[0067] Based on the above technical solution, by transforming the composite scheduling utility function into a cost matrix, the key problem of "incompatibility between optimization objectives and solution algorithms" in intelligent task allocation is effectively solved. In gig worker scheduling scenarios, matching needs to consider multiple factors such as skill, location, reputation, and fatigue, forming a complex nonlinear utility evaluation. However, mainstream efficient matching algorithms (such as the Hungarian algorithm) only support minimizing linear cost problems. If the original utility values ​​are used directly, mature solvers cannot be invoked; if infeasible pairings are ignored, invalid solutions may be output. This solution constructs the cost matrix by taking negative values ​​and setting infeasible edges to infinity. This preserves the utility ranking relationship and strictly transforms the problem into a standard minimum cost allocation problem, enabling the system to efficiently invoke classic combinatorial optimization algorithms and obtain the globally optimal feasible solution in polynomial time. This method combines theoretical rigor with engineering practicality, significantly improving the solution efficiency and result reliability of large-scale real-time scheduling.

[0068] S303. Using a variant of the Hungarian algorithm with dynamic pruning, solve the minimum-cost maximum-weight matching problem on the bipartite graph to obtain the result that maximizes total utility. Maximize the initial task allocation scheme .

[0069] The candidate edge set is the set of all task-person pairings that satisfy hard feasibility constraints (such as skill matching and time window reachability) and matching structure constraints (such as one-to-one and capacity limits) at the current scheduling time. In each iteration of the algorithm, before performing augmenting path search, the candidate edge set is dynamically pruned, i.e., edges that are no longer feasible due to environmental changes (such as offline personnel, traffic congestion, and task cancellation). The Hungarian algorithm variant solves the minimum cost maximum weight matching problem on the bipartite graph by maintaining a dynamically updated candidate edge set, and finally obtains the minimum cost maximum weight matching problem on the bipartite graph, maximizing the total utility. Maximize the initial task allocation scheme .

[0070] In some implementations, the dynamic pruning mechanism is triggered by real-time event streams. For example, when a person's status changes to "offline" or "timeout without order," all associated edges of that person are immediately removed from the candidate edge set. Matching structure constraints include at most one person assigned to each task and at most one order per person, which can be quickly verified through graph traversal or index tables. In a preferred implementation, the algorithm adopts an incremental update strategy, rebuilding the candidate edge set only when a new task is added or the person's status changes, avoiding full recalculation. Furthermore, it can be combined with a parallel computing framework to improve the solution efficiency in large-scale scenarios.

[0071] It should be noted that the above-mentioned algorithm variant does not directly run the traditional Hungarian algorithm. Instead, it introduces a dynamic pruning mechanism to transform the static graph optimization problem into a dynamic graph processing problem, which significantly improves the adaptability and real-time performance of the algorithm in high-frequency and highly dynamic scenarios. This method effectively avoids invalid searches for invalid edges and reduces computational complexity. At the same time, by preserving matching structure constraints, it ensures that the output results are valid matches, avoids abnormal situations such as "one person accepting multiple orders", and ensures the stability and compliance of the scheduling system.

[0072] Based on the above technical solutions, by introducing a dynamic pruning mechanism into the Hungarian algorithm, key problems such as "response lag, computational redundancy, and result invalidation" in high-dynamic gig worker scheduling scenarios of traditional static matching algorithms are effectively solved. Because the status of tasks and personnel changes in real time (e.g., personnel offline, sudden traffic changes, task cancellation), if the matching graph is constructed and solved all at once at the beginning of the scheduling cycle, the resulting solution may be partially invalid before execution, leading to allocation failure or frequent rescheduling. This solution dynamically updates the candidate edge set before each round of augmenting path search, eliminating infeasible pairings to ensure that the search is always based on the latest environmental state; at the same time, it retains matching structure constraints to ensure the legality of the solution. This method integrates classical combinatorial optimization algorithms with a real-time event-driven mechanism, significantly improving the robustness and real-time performance of the system without sacrificing optimality, providing an efficient and reliable solution foundation for large-scale, high-frequency intelligent scheduling.

[0073] In one possible implementation of the embodiments of this application, such as Figure 3 As shown, the above S204 can be specifically implemented through the following S401, S402, S403, S404 and S405, which are explained in detail below: The dynamic rematching mechanism includes the following matching process: S401, Construct the state space , is used to characterize the dynamic configuration of the current scheduling environment.

[0074] Each state Encodes the set of incomplete tasks Aggregator with available gig workers The combined information.

[0075] It should be noted that the state space The construction is not simply a matter of concatenating task and personnel sets into an input vector. Instead, it uses structured coding to incorporate the dynamic configuration information of the current scheduling environment, including the set of incomplete tasks. Aggregator with available gig workers The joint state is abstracted into a unified high-dimensional representation, thus providing a generalizable and perceptible decision-making basis for subsequent reinforcement learning strategies. The design of this state space fully considers the time-varying nature and complexity of the gig worker scheduling system: on the one hand, It includes multi-dimensional attributes such as task location, time window, and skill requirements, reflecting the system's pending workload; on the other hand... By integrating dynamic features such as personnel location, skill profiles, fatigue index, and historical reputation, the resource supply status can be reflected. This is achieved through joint encoding of these features (e.g., based on graph neural networks or attention mechanisms), resulting in a state space... This approach not only captures potential matching relationships between tasks and personnel but also supports modeling global resource allocation trends, laying the foundation for agents to make long-term optimal decisions in complex environments. This design avoids the problems of poor policy generalization and low training efficiency caused by incomplete state representation in traditional methods, and is a key prerequisite for achieving end-to-end dynamic rematching.

[0076] S402, Based on state space Constructing action space .

[0077] Each action This corresponds to making a new set of task-person pairing decisions between the currently unassigned or reassigned subset of tasks and the available personnel subset.

[0078] S403, Define the reward function Its value is in the state Next action The resulting increase in global utility.

[0079] The global utility increment is the sum of the overall scheduling benefits corresponding to all assigned tasks in the updated task allocation scheme.

[0080] It should be noted that the reward function The definition is not based on the immediate benefits of a single task, but rather on the global utility increment, that is, within the state. Next action Afterwards, the overall system scheduling benefit is significantly improved compared to before execution. This design overcomes the limitations of traditional reinforcement learning's "local rewards" or "rewards upon task completion," preventing agents from falling into short-sighted behavior (such as prioritizing high-value but resource-wasting tasks) and instead guiding them to focus on long-term system performance optimization. Specifically, global utility is the sum of the comprehensive scheduling benefits corresponding to all assigned tasks in the updated allocation scheme, encompassing multiple dimensions such as spatial accessibility, skill matching, reputation incentives, and fatigue suppression, ensuring that the reward signal truly reflects the comprehensive impact of the decision on the overall system efficiency, fairness, and sustainability. By introducing this incremental reward mechanism, the model can learn more strategic re-matching behaviors, such as proactively retaining highly capable personnel to handle future high-value tasks, or prioritizing critical orders when capacity is tight, thereby achieving an evolution from "order-by-order response" to "global collaborative optimization," significantly improving the intelligence and robustness of the dynamic scheduling system.

[0081] S404, Status The vectorized representation is then input into a deep Q-network model, which outputs the action space. The Q value corresponds to each action, where Q is the cumulative reward expected to be obtained by taking each rematching action in the current state.

[0082] Among them, state This represents a complete snapshot of the environment at the current scheduling moment, including the set of tasks to be assigned, the set of available gig workers, and their respective multidimensional attributes (such as location, skill profile, fatigue index, time window, etc.). This state is converted into a fixed-dimensional vector representation through a pre-defined encoder module. As input to the Deep Q-Network (DQN) model; action space Defined as the set of all possible rematch operations, such as "task". Reassigned to personnel "Or "Keep the current allocation unchanged"; the deep Q-network outputs each action. The corresponding Q value That is, in the state Next action The expected future discounts and cumulative rewards are used to guide strategy selection.

[0083] In some implementations, state vectorization is achieved using graph neural networks (GNNs) or Transformer encoders: tasks and people are treated as graph nodes, their attributes as node features, edges represent potential matching relationships, and global context information is aggregated through message passing mechanisms to generate compact and semantically rich state embeddings; deep Q-networks consist of multiple fully connected layers, with state vectors as input. The output dimension is equal to the size of the action space. In a preferred embodiment, to alleviate the problem of the action space exploding with the number of tasks / personnel, an action masking mechanism is adopted, outputting a valid Q value only for currently feasible rematching operations, and setting the remaining actions to negative infinity; in addition, model parameters Stable training is achieved through experience replay and target network mechanisms, and online fine-tuning is supported to adapt to changes in environmental distribution.

[0084] It should be noted that the aforementioned state vectorization and Q-value prediction processes are not isolated function mappings, but rather a key bridge that transforms complex combinatorial optimization problems into sequential decision-making problems in reinforcement learning: by compressing high-dimensional heterogeneous scheduling states into learnable vector representations, DQN can capture the implicit dependencies between tasks and personnel and the global resource situation; while the "expected cumulative reward" represented by the Q-value naturally integrates short-term utility and long-term benefits, enabling the agent to balance immediate allocation quality and future scheduling flexibility in dynamic rematching; more importantly, this method avoids the dependence of traditional rule-based or heuristic algorithms on manually designed strategies, and achieves data-driven adaptive optimization capabilities.

[0085] S405. Each time dynamic rematching is triggered, the action with the largest Q value is selected as the optimal rematching strategy based on the Q value output by the deep Q-network model, and an updated task allocation scheme is generated based on this action. .

[0086] It should be noted that the Q-value used is not a static rule or heuristic scoring result, but a dynamic estimate of future cumulative utility learned by a deep Q-network model based on a large amount of historical scheduling interaction experience; this step, by selecting the action with the maximum Q-value, is essentially based on the current state. The approach employs a near-optimal strategy, thereby transforming the complex combinatorial rematching problem into a computable decision process. Furthermore, the generated updated task allocation scheme... It not only reflects the quality of real-time matching, but also implies forward-looking considerations for subsequent scheduling flexibility and system load balancing.

[0087] The methods for iteratively optimizing the Q-value of deep Q-network models include: ; in, This is the current state. To perform the action The new state afterwards; For the new state Candidate actions to be executed, Main network parameters, For target network parameters; For learning rate, Discount factor; Indicates the state Take action below The expected cumulative reward.

[0088] The above Q-value calculation is applied to obtain the expected cumulative reward. The cumulative reward is then weighted, summed, and averaged to obtain the iteratively optimized Q-value.

[0089] Among them, through the temporal difference (TD) learning mechanism, the agent can gradually approach the optimal decision strategy from practical experience in a dynamic scheduling environment; This constitutes an estimate of the "real long-term return" (target value), minus the current forecast. Obtain the error signal, and then use the learning rate Controlling the update step size achieves stable convergence; simultaneously, target network parameters are introduced. Decoupling prediction from the objective effectively suppresses training oscillations. This not only enables the system to balance immediate utility and future gains, but also ensures that deep Q-networks can efficiently and stably learn forward-looking and robust rematching strategies in high-dimensional, nonlinear, and highly dynamic gig work scheduling scenarios.

[0090] It should be noted that the Q-value mentioned above is not manually set or calculated using simple weighted rules, but rather a policy estimate output by the deep Q-network after automatically modeling the complex nonlinear relationships between tasks, personnel, and the environment by learning from massive scheduling experience. This value not only reflects the immediate benefits of the current action, but also contains the ability to predict future state evolution and long-term system performance. Therefore, selecting an action based on the maximum Q-value is essentially executing a near-optimal decision in a dynamic environment, realizing an intelligent leap from "rule-driven" to "data-driven". At the same time, the continuous iterative optimization of the Q-value ensures that the model can adapt to business changes and environmental disturbances, and has good generalization and adaptive capabilities.

[0091] Some of the data in the above formula are calculated by removing dimensions and taking their numerical values. The formula is the closest to the real situation obtained by software simulation of a large amount of collected data. The preset parameters and preset thresholds in the formula are set by those skilled in the art according to the actual situation or obtained through simulation of a large amount of data.

Claims

1. A task scheduling method for gig workers based on artificial intelligence, characterized in that, include: Obtain information on tasks to be scheduled and information on gig workers, and construct a composite scheduling utility function; The comprehensive scheduling benefits of allocating tasks to gig workers are evaluated based on a composite scheduling utility function. Using the comprehensive scheduling benefits as matching weights, a task-person bipartite graph is constructed. Under the conditions of satisfying hard feasibility constraints and matching structure constraints, the maximum weight matching problem of the bipartite graph is solved to obtain the initial task allocation scheme. Among them, the hard feasibility constraints include task time window constraints and personnel reachability constraints, and the matching structure constraints include that each task can be assigned to at most one gig worker and each gig worker can undertake at most one task. Task execution is initiated based on the initial task allocation scheme, and the environmental status is monitored in real time during execution. If the environmental status is abnormal, a dynamic re-matching mechanism is initiated with the current set of unfinished tasks and the status of temporary workers as input to generate an updated task allocation scheme. Among them, abnormal environmental status includes any assigned task that cannot be completed as originally planned due to personnel withdrawal, traffic abnormalities, task changes, or timeout risks.

2. The method for scheduling gig workers based on artificial intelligence according to claim 1, characterized in that, The method for constructing the composite scheduling utility function includes: The task information to be scheduled includes the task's geographical location. Skills demand vector Time window and urgency weight ;in, Let i be the i-th task to be scheduled, where i is the index number of the task to be scheduled. Tasks to be scheduled Early start time, Tasks to be scheduled Late end time; The personnel information includes the real-time location of casual workers. Capability profile vector Current fatigue index and historical performance rating ;in, Let j be the j-th gig worker, where j is the index number of the gig worker; The composite scheduling utility function Defined as: ； in, These are preset parameters for the spatial attenuation scale. This is a skill-dimensional covariance matrix estimated based on historical successful matching data, used to characterize the correlation and tolerance resilience among different skills. The total number of skill dimensions. It is the inverse of the covariance matrix of the skill dimension. To balance the weighting of skill matching and historical reputation, This is a parameter for adjusting fatigue sensitivity.

3. The method for scheduling gig workers based on artificial intelligence according to claim 2, characterized in that, The Capability Profile Methods for obtaining [the information] include: Extracting gig workers by analyzing platform user registration information and historical task completion records. Behavioral evidence across multiple preset skill dimensions, including professional qualification certificates, frequency of task type completion, user evaluation keywords, and platform certification level; Constructing a task-skill mapping matrix ;in, The total number of task categories. The total number of skill dimensions. Indicates the first Class of tasks requires the first Item skill; For each skill dimension Calculate the number of gig workers In the Overall ability score in each skill: ； in, Personnel Whether to hold the execution of the Minimum qualification certificate required for this skill; For personnel Complete the first within the historically preset timeframe. Total number of tasks of this type; This serves as an index when iterating through or taking extreme values ​​of all gig workers. To extract information about the first point from user comments based on natural language processing Emotional score for each skill; Preset weighting coefficients; This is the Sigmoid normalization function.

4. The method for scheduling gig workers based on artificial intelligence according to claim 2, characterized in that, The fatigue index Methods for obtaining [the information] include: Real-time data collection of gig workers Terminal sensor data and task execution logs, including the continuous working hours for the day. Number of tasks completed per unit of time Movement speed volatility and physiological indicators of heart rate variability; The sensor data and task execution logs are input into a pre-trained fatigue perception neural network model, which outputs the current fatigue index. The fatigue perception neural network model is a three-layer fully connected network, trained on a labeled dataset. The labels are the fatigue levels assessed by experts based on task timeout rate, frequency of operational errors, and overall evaluation.

5. The method for scheduling gig workers based on artificial intelligence according to claim 2, characterized in that, The historical performance score Methods for obtaining [the information] include: Based on gig workers Historical task performance records, calculating historical performance scores. : ； in, The number of historical tasks considered in the current time-sharing calculation of contract fulfillment. This serves as a sequence number index for historical tasks. As an indicator variable for whether it is completed on time, Rate the task for the user. For time decay weight, For the first The time of completion of this task. To adjust the parameters.

6. The method for scheduling gig workers based on artificial intelligence according to claim 2, characterized in that, The process of constructing the task-person bipartite graph includes: Based on the aforementioned comprehensive scheduling benefit value As edge weights, construct a task-person bipartite graph. The vertex set consists of the set of tasks to be scheduled. Gathering with casual workers constitute, The total number of tasks to be scheduled. The total number of gig workers, side collection This includes all task-person pairings that meet the hard feasibility constraints, and each edge Weighted ; The hard feasibility constraints include: Personnel accessibility constraints: gig workers Capability Profile Vector In the mission The projection on the required skill dimension is not lower than the preset threshold; Task time window constraint: Travel time calculated based on a real-time traffic flow prediction model satisfy: ； Where t is the current time, The average working speed, The minimum duration required for the task to be executed.

7. The method for scheduling gig workers based on artificial intelligence according to claim 6, characterized in that, The method for obtaining the initial matching scheme includes: Based on the bipartite graph Construct the cost matrix For each feasible edge Its corresponding element is defined as ; A variant of the Hungarian algorithm with dynamic pruning is used to solve the minimum-cost maximum-weight matching problem on this bipartite graph, obtaining the result that maximizes total utility. Maximize the initial task allocation scheme In each iteration of the algorithm, before performing augmenting path search, the candidate edge set is dynamically pruned. The candidate edge set is the set of all task-person pairings that satisfy the hard feasibility constraints and matching structure constraints at the current scheduling time.

8. The method for scheduling gig workers based on artificial intelligence according to claim 1, characterized in that, The dynamic rematching mechanism includes the following matching process: Constructing the state space This is used to characterize the dynamic configuration of the current scheduling environment; where each state... Encodes the set of incomplete tasks Aggregator with available gig workers Joint information; Based on the state space Constructing action space ; among them, each action This corresponds to making a new set of task-person pairing decisions between the currently unassigned or reassigned subset of tasks and the available personnel subset; Define reward function Its value is in the state Next action The resulting global utility increment is the sum of the comprehensive scheduling benefits corresponding to all assigned tasks in the updated task allocation scheme; The state The vectorized representation is then input into a deep Q-network model, which outputs the action space. The Q value corresponds to each action, where Q is the cumulative reward expected to be obtained by taking each rematching action in the current state. Each time dynamic rematch is triggered, the action with the largest Q value is selected as the optimal rematch strategy based on the Q value output by the deep Q-network model, and an updated task allocation scheme is generated based on this action. .

9. A method for scheduling gig workers based on artificial intelligence according to claim 8, characterized in that, The method for iteratively optimizing the Q-value of the deep Q-network model includes: ； in, This is the current state. To perform the action The new state afterwards; For the new state Candidate actions to be executed, Main network parameters, For target network parameters; For learning rate, Discount factor; Indicates the state Take action below The expected cumulative reward.

10. An AI-based gig worker task scheduling system, operating based on the AI-based gig worker task scheduling method according to any one of claims 1-9, characterized in that, It includes an analysis module, an allocation module, and an execution module; The analysis module is used to obtain information on tasks to be scheduled and information on gig workers, and to construct a composite scheduling utility function; based on the composite scheduling utility function, it evaluates the comprehensive scheduling benefits of allocating tasks to gig workers. The allocation module is used to construct a task-personnel bipartite graph using the comprehensive scheduling benefit as the matching weight, and solve the maximum weight matching problem of the bipartite graph under the conditions of satisfying hard feasibility constraints and matching structure constraints to obtain an initial task allocation scheme. The execution module is used to start task execution based on the initial task allocation scheme and monitor the environmental status in real time during execution. If the environmental status is abnormal, the dynamic re-matching mechanism is started with the current set of unfinished tasks and the status of temporary workers as input to generate an updated task allocation scheme.

Images