Intelligent scheduling control method in complex environment based on imitation learning

Abstract

Embodiments of the present disclosure disclose an intelligent scheduling control method in a complex environment based on imitation learning. The specific implementation of the method comprises: calling a large language model to analyze text and convert it into a structured data set, and switching to a conditional random field model when it fails; calculating the minimum Euclidean distance between the environment context vector and the historical expert trajectory matrix to establish the state trajectory envelope drift distance, and when it is greater than a threshold, reconstructing the local expert trajectory segment and refreshing the weight; projecting the policy update gradient to the orthogonal complement space of the expert behavior cloning gradient, calculating the actual update gradient and refreshing the policy parameters to output instructions. This implementation suppresses parameter oscillation caused by state deviation, restricts the update direction using gradient space orthogonality, avoids weight rewriting, and ensures stable scheduling control flow.

Description

Technical Field

[0001] The embodiments disclosed herein relate to the field of automated scheduling and control technology, and more specifically to an intelligent scheduling and control method based on imitation learning in complex environments. Background Technology

[0002] Currently, in highly dynamic resource allocation and intellectual activity management scenarios, the mainstream control strategy adopts a control method based on state-space adaptive search and reward function feedback guidance. This method uses deep networks to approximate the mapping function between complex operating conditions and controlled variables, and adjusts the control network weights through continuous interactive trial and error with non-ideal environments. Internally, the system adjusts control parameters through time-series differential updates to maintain the convergence of the state-value function. This technical approach can provide basic scheduling control flow under standard operating conditions.

[0003] However, scheduling tasks exhibit high-frequency instantaneous fluctuations. When the system faces multiple shocks such as task overlap and heterogeneous attribute boundary constraints, the dimensionality of the state space expands. Due to the sparse effective feedback signals, the policy network is prone to blind random exploration. Disturbances generated by online conditions also cause covariate shifts in the environmental context vector. This change causes the current state features to slip out of the distribution support set of the initial stable policy trajectory, resulting in high-frequency oscillations in decision parameters and leading to the break of the global scheduling link. The industry usually smooths the data distribution by expanding the capacity of the historical experience replay cache or by using manual intervention to reconstruct the weighting of reward factors. However, such improvement methods are prone to introducing new policy distribution distortions, leading to increased computational latency. This approach cannot reconcile the conflict between the ability to broaden the online boundary and the stability of existing control experience. Although some solutions improve from the perspective of physical layout or hardware structure limitations, the existing technology also has shortcomings at the core software control method level. For example, Chinese invention patent application CN122114530A discloses a method and device for real-time power grid dispatch decision-making using large-model collaborative reinforcement learning. This method parses text using a large language model to generate semantic feature vectors, which are then concatenated with state vectors and input into the reinforcement learning agent. However, this approach relies on a stable state space distribution. When faced with extreme conditions involving high-frequency instantaneous fluctuations and intertwined multidimensional hard constraints, the environmental context vectors exhibit severe covariate shifts. Furthermore, the lack of real-time calibration capability for state trajectory envelopes means that direct feature concatenation and conventional gradient updates can easily lead to conflicts between the online reinforcement learning gradient and the existing expert experience gradient direction, causing high-frequency oscillations in decision parameters and resulting in overwriting of existing weight distributions under normal operating conditions and catastrophic forgetting.

[0004] Therefore, the technical problem to be solved by this invention is how to overcome the state trajectory deviation in online random exploration, complete the real-time closed-loop self-calibration of the controlled state trajectory envelope under multidimensional constraints, and constrain the evolution direction of the decision strategy during mixed data updates to prevent the weight parameter distribution from being overwritten. Summary of the Invention

[0005] The summary portion of this disclosure is intended to provide a brief overview of the concepts, which will be described in detail in the detailed description portion. This summary portion is not intended to identify key or essential features of the claimed technical solutions, nor is it intended to limit the scope of the claimed technical solutions.

[0006] Some embodiments of this disclosure propose an intelligent scheduling and control method based on imitation learning in complex environments to solve the technical problems mentioned in the background section above.

[0007] In a first aspect, some embodiments of this disclosure provide an intelligent scheduling and control method for complex environments based on imitation learning, the method comprising: Step S101: Call the large language model to calculate the internal self-attention weight matrix of the input unstructured task requirement text, parse the implicit constraint information and extract the task attribute fields, convert it into a structured dataset and output it to the system data bus, synchronously call the compliance detection unit to determine the integrity of the structured dataset, and trigger the circuit breaker mechanism when encountering data distortion or failure of the large language model, and switch to the conditional random field model to obtain environmental scheduling requirement data. Step S102: Before forward propagation calculation, the processor reads the current environment context vector and retrieves the historical expert demonstration trajectory matrix. It then iterates through and calculates the minimum Euclidean distance between each trajectory vector in the environment context vector and the historical expert demonstration trajectory matrix to establish the state trajectory envelope drift distance. Step S103: When the state trajectory envelope drift distance is greater than the preset boundary threshold, suspend the current forward computation flow, route the state trajectory envelope drift distance as a penalty to the multi-objective optimization layer, increase the mutation probability threshold of the non-dominated sorting genetic algorithm and proportionally increase the local adjustment step size of the feasible solution repair operator, reconstruct the local expert trajectory segment with the current state as the center and refresh the network weights in situ. Step S104: Calculate the expert behavior clone gradient vector based on the local expert trajectory segment, synchronously read the current policy update gradient vector corresponding to the time-series difference error, project the current policy update gradient vector onto the orthogonal complement space of the expert behavior clone gradient vector, calculate and generate the actual update gradient vector, and refresh the policy network weight parameters in situ to output dynamic scheduling control instructions.

[0008] Preferably, the task attribute fields in step S101 include task type, physical spatial location coordinates, time window limits, skill qualification requirements, and basic remuneration standards; the structured dataset is a structured JSON dataset.

[0009] Preferably, step S101 includes the following sub-steps: Step S1011: Use a large language model to identify the relationship between entity labels and physical constraints in unstructured task requirement text, and map them to generate a structured JSON dataset containing task requirement dimensions, spatial distribution coordinates, and individual attributes. In step S1012, when the integrity of the structured JSON dataset is determined to be incomplete, the compliance detection unit triggers the circuit breaker mechanism, shuts down the large language model and switches to the conditional random field model, and identifies sequence labels through time series conditional probability statistics to complete the environmental scheduling requirement data.

[0010] Preferably, step S102 includes the following sub-steps: Step S1021: Read the current environment scheduling requirement data from the system data bus and extract feature indicators including the number of tasks, personnel distribution and collaborative boundary constraints. Map the feature indicators into high-dimensional feature vectors through a fully connected layer to serve as the current environment context vector. Step S1022: Retrieve the offline-generated expert optimization scheduling scheme sequence from the historical demonstration database and convert the expert optimization scheduling scheme sequence into the corresponding state action trajectory vector, and perform sub-storage block partitioning to form a historical expert demonstration trajectory matrix containing multiple sets of time-series associated states.

[0011] Preferably, in step S103, increasing the mutation probability threshold of the non-dominated sorting genetic algorithm specifically involves summing the preset baseline mutation probability value with the product of the adjustment coefficient and the state trajectory envelope drift distance to update the mutation probability threshold of the non-dominated sorting genetic algorithm.

[0012] Preferably, step S103 includes the following sub-steps: Step S1031: Increase the local adjustment step size of the feasible solution repair operator, and use the feasible solution repair operator to map and correct the infeasible solution space that violates the spatial distribution constraints of the supply side and the heterogeneous individual attribute restrictions, so as to generate a candidate solution set that satisfies the boundary hard constraints. Step S1032: The dynamic requirements extracted by the large language model are used as the current state center. The multi-objective optimization algorithm is used to search for the optimal solution with the smallest difference from the historical expert demonstration trajectory matrix in the candidate solution set. The local expert trajectory segment is reconstructed and the network weights are refreshed in situ.

[0013] Preferably, step S104 includes the following sub-steps: Step S1041: Construct a behavior cloning loss function containing the first derivative term with respect to the policy network weight parameters, and calculate the first derivative of the behavior cloning loss function with respect to the policy network weight parameters based on the local expert trajectory fragments to generate the expert behavior cloning gradient vector. Step S1042: Synchronously read the timing difference error in the current control cycle, calculate the gradient of the timing difference error with respect to the value function to generate the current policy update gradient vector; Step S1043: Construct the orthogonal complement space of the expert behavior clone gradient vector, project the current policy update gradient vector onto the orthogonal complement space of the expert behavior clone gradient vector to obtain orthogonal gradient components, use the orthogonal gradient components as the actual update gradient vector and refresh the policy network weight parameters in situ, and restrict the online update direction to be in the same direction as the expert behavior clone gradient vector.

[0014] Preferably, the process of updating the policy network weight parameters also includes the following steps: Step S105: Establish an experience replay buffer to store the state, actions, reward feedback signals, and subsequent state data during the historical scheduling process; Step S106: Randomly extract a small batch of sample data from the experience replay buffer to decouple the temporal correlation of historical data; Step S107: Simultaneously construct the target value network and use the target value network to calculate the target value function value in order to maintain the stability of the value function update.

[0015] Preferably, the reward feedback signal in step S105 is generated through the following sub-steps: Step S1051: Monitor the task completion quality of supply-side personnel and the service satisfaction of demand-side personnel in real time, and calculate the positive revenue reward value. Step S1052: Calculate the geographic distance loss and individual collaboration cost of heterogeneous personnel in the scheduling process, and generate a negative penalty term; Step S1053: The positive reward value and the negative penalty term are weighted and summed to output a reward feedback signal used to guide the update of the policy network weight parameters.

[0016] Preferably, the output of the dynamic scheduling control command in step S104 includes the following sub-steps: Step S1044: Apply the in-situ refreshed policy network weight parameters to the forward propagation calculation and output an adaptive real-time dynamic scheduling control matrix. Step S1045: Establish a two-dimensional row and column structure of the real-time dynamic scheduling control matrix, where the number of rows corresponds to the total number of tasks to be assigned and the number of columns corresponds to the total number of personnel that can be deployed. Each cross element in the real-time dynamic scheduling control matrix represents the joint scheduling matching probability value between the corresponding task and the corresponding personnel. Based on the joint scheduling matching probability value, issue personnel dispatch control instructions to output dynamic scheduling control instructions.

[0017] The above-described embodiments of this disclosure have the following beneficial effects: 1. In intelligent scheduling and control under complex environments, an offline expert demonstration trajectory is generated using a multi-objective optimization algorithm. Combined with behavioral cloning supervision signals, the cross-entropy between the policy network output and the trajectory is minimized. This internalizes the offline expert scheduling experience into the network's initial weight parameters. During crossover and mutation evolution, a feasible solution repair operator is used to reverse-search and replace conflicting gene positions, forcibly correcting the daily work hour upper limit constraint, enabling subsequent policy gradient optimization to bypass the stochastic exploration phase. This offline constraint repair and supervised initialization mechanism works together to avoid the decision network blindly trying and failing in the initial high-dimensional state space, stabilizes the variance of the network value function estimation, and improves the initial convergence speed of weights in complex, multi-constraint environments.

[0018] 2. Before forward propagation, the central processing unit (CPU) calculates the minimum Euclidean distance between the current environment context vector and the historical expert trajectory vectors in the cache block to establish the state trajectory envelope drift distance. When the drift distance exceeds a preset boundary threshold, the current forward computation flow is suspended, and the drift distance is used as a penalty for reverse routing to the multi-objective optimization layer. The mutation probability of the genetic algorithm is adaptively increased, and the adjustment step size of the feasible solution repair operator is amplified, enabling the system to reconstruct local expert trajectory segments centered on the current state and refresh the network weights in situ. This real-time state tracking and feedback fine-tuning loop forms a dependent closed loop between the preceding and following steps, suppressing the oscillation of policy network parameters caused by online state shifts.

[0019] 3. The central processing unit (CPU) calculates the expert behavior clone gradient vector based on local expert trajectory fragments, synchronously reads the current policy update gradient vector corresponding to the time-series difference error, and projects the current policy update gradient vector onto the orthogonal complement space of the expert behavior clone gradient vector. This generates the actual update gradient vector and refreshes the policy network weight parameters in situ. This orthogonal constraint of the gradient space restricts the update direction of the online policy to remain orthogonal or in the same direction as the gradient direction of the basic policy. While adapting to delay-free control reorganization under unknown edge disturbances, it locks in the existing weight distribution benchmark under standard operating conditions, effectively avoiding catastrophic forgetting caused by heterogeneous data mixing updates. Attached Figure Description

[0020] The above and other features, advantages, and aspects of the embodiments of this disclosure will become more apparent from the accompanying drawings and the following detailed description. Throughout the drawings, the same or similar reference numerals denote the same or similar elements. It should be understood that the drawings are schematic, and elements are not necessarily drawn to scale.

[0021] Figure 1 This is a flowchart of the intelligent scheduling and control method based on imitation learning of the present invention; Figure 2 This is a diagram illustrating the environmental scheduling requirements for the large language model failure circuit breaker in this invention. Figure 3This is a diagram showing the actual updated gradient vector calculation for the orthogonal space projection of this invention. Detailed Implementation

[0022] Embodiments of this disclosure will now be described in more detail with reference to the accompanying drawings. While some embodiments of this disclosure are shown in the drawings, it should be understood that this disclosure can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of this disclosure. It should be understood that the accompanying drawings and embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of protection of this disclosure.

[0023] It should also be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings. Unless otherwise specified, the embodiments and features described in this disclosure can be combined with each other.

[0024] It should be noted that the concepts of "first" and "second" mentioned in this disclosure are used only to distinguish different devices, modules or units, and are not used to limit the order of functions performed by these devices, modules or units or their interdependencies.

[0025] It should be noted that the terms "a" and "a plurality of" used in this disclosure are illustrative rather than restrictive, and those skilled in the art should understand that, unless otherwise expressly indicated in the context, they should be understood as "one or more".

[0026] The names of messages or information exchanged between multiple devices in the embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of such messages or information.

[0027] Before performing any of the operations involving the collection, storage, processing, or use of the target images disclosed herein, the relevant organizations or individuals shall fulfill their obligations, including conducting information security impact assessments, informing the information subjects, and obtaining prior authorization and consent from the information subjects.

[0028] This disclosure will now be described in detail with reference to the accompanying drawings and embodiments.

[0029] refer to Figure 1 This document illustrates the flowchart of some embodiments of an intelligent scheduling and control method for complex environments based on imitation learning, according to the present disclosure. The intelligent scheduling and control method for complex environments based on imitation learning includes the following steps: Step S101: Call the large language model to calculate the internal self-attention weight matrix of the input unstructured task requirement text, parse the implicit constraint information and extract the task attribute fields, convert it into a structured dataset and output it to the system data bus, synchronously call the compliance detection unit to determine the integrity of the structured dataset, and trigger the circuit breaker mechanism when encountering data distortion or failure of the large language model, and switch to the conditional random field model to obtain environmental scheduling requirement data. Step S102: Before forward propagation calculation, the processor reads the current environment context vector and retrieves the historical expert demonstration trajectory matrix. It then iterates through and calculates the minimum Euclidean distance between each trajectory vector in the environment context vector and the historical expert demonstration trajectory matrix to establish the state trajectory envelope drift distance. Step S103: When the state trajectory envelope drift distance is greater than the preset boundary threshold, suspend the current forward computation flow, route the state trajectory envelope drift distance as a penalty to the multi-objective optimization layer, increase the mutation probability threshold of the non-dominated sorting genetic algorithm and proportionally increase the local adjustment step size of the feasible solution repair operator, reconstruct the local expert trajectory segment with the current state as the center and refresh the network weights in situ. Step S104: Calculate the expert behavior clone gradient vector based on the local expert trajectory segment, synchronously read the current policy update gradient vector corresponding to the time-series difference error, project the current policy update gradient vector onto the orthogonal complement space of the expert behavior clone gradient vector, calculate and generate the actual update gradient vector, and refresh the policy network weight parameters in situ to output dynamic scheduling control instructions.

[0030] Preferably, the task attribute fields in step S101 include task type, physical spatial location coordinates, time window limits, skill qualification requirements, and basic remuneration standards; the structured dataset is a structured JSON dataset.

[0031] Preferably, step S101 includes the following sub-steps: Step S1011: Use a large language model to identify the relationship between entity labels and physical constraints in unstructured task requirement text, and map them to generate a structured JSON dataset containing task requirement dimensions, spatial distribution coordinates, and individual attributes. In step S1012, when the integrity of the structured JSON dataset is determined to be incomplete, the compliance detection unit triggers the circuit breaker mechanism, shuts down the large language model and switches to the conditional random field model, and identifies sequence labels through time series conditional probability statistics to complete the environmental scheduling requirement data.

[0032] Preferably, step S102 includes the following sub-steps: Step S1021: Read the current environment scheduling requirement data from the system data bus and extract feature indicators including the number of tasks, personnel distribution, and collaborative boundary constraints. Map the feature indicators into high-dimensional feature vectors through a fully connected layer to serve as the current environment context vector; S1022, retrieve the offline-generated expert optimization scheduling scheme sequence from the historical demonstration database and convert the expert optimization scheduling scheme sequence into the corresponding state action trajectory vector, and perform sub-storage block partitioning to form a historical expert demonstration trajectory matrix containing multiple sets of time-series associated states.

[0033] Preferably, in step S103, increasing the mutation probability threshold of the non-dominated sorting genetic algorithm specifically involves summing the preset baseline mutation probability value with the product of the adjustment coefficient and the state trajectory envelope drift distance to update the mutation probability threshold of the non-dominated sorting genetic algorithm.

[0034] Preferably, step S103 includes the following sub-steps: Step S1031: Increase the local adjustment step size of the feasible solution repair operator, and use the feasible solution repair operator to map and correct the infeasible solution space that violates the spatial distribution constraints of the supply side and the heterogeneous individual attribute restrictions, so as to generate a candidate solution set that satisfies the boundary hard constraints. Step S1032: The dynamic requirements extracted by the large language model are used as the current state center. The multi-objective optimization algorithm is used to search for the optimal solution with the smallest difference from the historical expert demonstration trajectory matrix in the candidate solution set. The local expert trajectory segment is reconstructed and the network weights are refreshed in situ.

[0035] Preferably, step S104 includes the following sub-steps: Step S1041: Construct a behavior cloning loss function containing the first derivative term with respect to the policy network weight parameters, and calculate the first derivative of the behavior cloning loss function with respect to the policy network weight parameters based on the local expert trajectory fragments to generate the expert behavior cloning gradient vector. Step S1042: Synchronously read the timing difference error in the current control cycle, calculate the gradient of the timing difference error with respect to the value function to generate the current policy update gradient vector; Step S1043: Construct the orthogonal complement space of the expert behavior clone gradient vector, project the current policy update gradient vector onto the orthogonal complement space of the expert behavior clone gradient vector to obtain orthogonal gradient components, use the orthogonal gradient components as the actual update gradient vector and refresh the policy network weight parameters in situ, and restrict the online update direction to be in the same direction as the expert behavior clone gradient vector.

[0036] Preferably, the process of updating the policy network weight parameters also includes the following steps: Step S105, establishing an experience replay buffer to store the state, actions, reward feedback signals and subsequent state data in the historical scheduling process; Step S106: Randomly extract a small batch of sample data from the experience replay buffer to decouple the temporal correlation of historical data; Step S107: Simultaneously construct the target value network and use the target value network to calculate the target value function value in order to maintain the stability of the value function update.

[0037] Preferably, the reward feedback signal in step S105 is generated through the following sub-steps: Step S1051: Monitor the task completion quality of supply-side personnel and the service satisfaction of demand-side personnel in real time, and calculate the positive revenue reward value. Step S1052: Calculate the geographic distance loss and individual collaboration cost of heterogeneous personnel in the scheduling process, and generate a negative penalty term; Step S1053: The positive reward value and the negative penalty term are weighted and summed to output a reward feedback signal used to guide the update of the policy network weight parameters.

[0038] Preferably, the output of the dynamic scheduling control instruction in step S104 includes the following sub-steps: Step S1044, applying the in-situ refreshed policy network weight parameters to the forward propagation calculation, and outputting an adaptive real-time dynamic scheduling control matrix. Step S1045: Establish a two-dimensional row and column structure of the real-time dynamic scheduling control matrix, where the number of rows corresponds to the total number of tasks to be assigned and the number of columns corresponds to the total number of personnel that can be deployed. Each cross element in the real-time dynamic scheduling control matrix represents the joint scheduling matching probability value between the corresponding task and the corresponding personnel. Based on the joint scheduling matching probability value, issue personnel dispatch control instructions to output dynamic scheduling control instructions.

[0039] Example 1: In a continuously running distributed intelligent computing service and cross-regional personnel scheduling management environment, computing devices access large-scale unstructured task requirement texts and historical scheduling evaluation records of flexible workers on the supply side in real time from the system data bus. However, due to high-frequency instantaneous fluctuations in demand and highly discrete spatial distribution and heterogeneous individual attribute constraints on the supply side, traditional static rule-based scheduling methods suffer from covariate shifts in the online-collected environmental context vectors when facing extreme conditions such as personnel canceling orders due to unforeseen circumstances or sudden emergency orders. This causes state features to slip out of the distribution support set of the initialization strategy trajectory, resulting in blind random exploration of the reinforcement learning control strategy in an uncovered high-dimensional state space. Random exploration leads to a sharp increase in the variance of the value function estimation and high-frequency oscillations in decision parameters. The system cannot converge to an effective configuration solution within a finite clock cycle, resulting in technical problems such as global control link blockage and large scheduling response delays.

[0040] The intelligent scheduling and control method of this invention operates as follows in a distributed intelligent computing service and cross-regional personnel scheduling and management environment: The central processing unit of the computing device accesses the unstructured task requirement text from the application layer. The processor calls a pre-trained large language model to calculate the internal self-attention weight matrix of the input unstructured task requirement text, parses the implicit constraint information, and extracts task attribute fields containing task type, physical spatial location coordinates, time window limits, skill qualification requirements, and basic reward standards. This is then converted into a structured dataset and output to the system data bus.

[0041] Specifically, the parsing and transformation process for extracting task attribute fields is implemented by externally connecting a lightweight linear decoding network after the last self-attention module of the large language model. The internal self-attention weight matrix output by the large language model is represented as a high-dimensional tensor under a multi-head attention mechanism. This invention reduces the dimensionality of this tensor by mean pooling along the attention head dimension and converts it into a bidirectional dependency weight feature matrix representing the relevance of words. This feature matrix is ​​then input into a pre-trained bidirectional long short-term memory network layer. This layer, combined with a pre-defined named entity recognition classifier, performs semantic alignment and boundary regression on specific strongly activated regions in the high-dimensional spatial features, directly identifying key entity character slices in the text stream that correspond to geofences, timestamp boundaries, skill levels, basic salary standards, etc. Finally, through a pre-defined structured mapping dictionary within the system, the extracted entity characters are filled into the structured data message in the form of key-value pairs, thereby completing the smooth conversion to a structured dataset that meets the reading protocol of the system data bus.

[0042] Simultaneously, the system calls the compliance detection unit to determine the integrity of the structured dataset. In the event of data distortion or failure of the large language model, it automatically triggers a circuit breaker strategy, switching to a conditional random field model to identify sequence labels through time-series conditional probability statistics to obtain environmental scheduling requirement data. Based on this, the system collects self-description texts written by personnel and enterprise evaluation texts, extracts structured skill tags using the large language model, and generates a structured report containing a comprehensive ability assessment. Embedding technology is used to map discrete text categories into continuous high-dimensional vectors, assembling and constructing a five-dimensional dynamic profile vector stored in continuous data blocks in system memory. The formula for the five-dimensional dynamic profile vector is as follows: V=[S,C,A,P,G], where V is the five-dimensional structure feature vector, S is the skill indicator, C is the reputation indicator, A is the availability indicator, P is the preference indicator, and G is the growth indicator. The value range of each dimension indicator is normalized and converges to a value between 0 and 1.

[0043] The transformation and dimensionality reduction from high-dimensional continuous embedding vectors to these five specific index scores are achieved through a pre-trained multi-label dense projection network layer. The continuous high-dimensional vectors generated by the large language model contain 512 dimensions. To map this high-dimensional vector to a unified engineering quantization channel, this invention inputs the high-dimensional vector into five parallel fully connected compression branches. The first compression branch calculates the co-occurrence probability of skill entity tags in the historical graph and outputs a scalar value representing the skill index. The second compression branch calculates a weighted average of the positive and negative sentiment polarity probabilities in the evaluation text and outputs a scalar value representing the reputation index. The third compression branch retrieves the variance of the historical check-in time series and outputs a scalar value representing the availability index. The fourth compression branch uses cosine similarity to measure the distance between the text embedding and the preset preference cluster center and outputs a scalar value representing the preference index. The fifth compression branch outputs a scalar value representing the growth index based on the monotonically increasing slope of the personnel's historical service level. The output of each of the five branches is connected to a logistic regression activation function, which independently normalizes and maps the five calculated scalar values ​​to a closed real number interval between 0 and 1, thereby assembling and constructing a complete five-dimensional structural feature vector.

[0044] Next, the system calls a deep collaborative filtering model to calculate the cosine similarity between the five-dimensional dynamic profile vector and the structured task feature vector. The calculated matching score is then used as the basic input parameter for subsequent multi-objective optimization and fed into a dedicated computing unit to initiate the non-dominated sorting genetic algorithm. This algorithm uses integer sequence encoding that satisfies hard constraints as the gene chromosome representation, where the position index in the sequence corresponds to a specific task node. The value at that position corresponds to the assigned agent number. During crossover and mutation evolution, the system activates a feasible solution repair operator. After the mutation operation, the cumulative working hours of a certain agent within the current time window satisfy the inequality: T i >Tmax , among which, T i T represents the cumulative working hours of agent i within the current time window. max The preset daily cumulative working hours limit is set to 8 hours. When the cumulative working hours satisfy this inequality, the system triggers the repair operator. The system searches for conflicting gene positions in reverse along the chromosome sequence and forcibly replaces their values ​​with values ​​that satisfy working hours less than or equal to T within the current time window. max Furthermore, the available agent IDs with the second-highest matching degree are assigned until the chromosome completely satisfies the constraints. The optimal solution set is then output as the historical expert demonstration trajectory matrix and written into a temporary matrix. This matrix serves as the supervisory signal for behavior cloning and is input into the scheduling control policy network constructed by the deep neural network. The system uses maximum likelihood estimation as the loss function to minimize the cross-entropy between the policy network's predicted output and the expert demonstration trajectory, thus converting offline expert scheduling experience into the initial weight parameters of the policy network.

[0045] Before the forward propagation computation of the system, the central processing unit (CPU) reads the current environment context vector and retrieves the historical expert demonstration trajectory matrix pre-stored in the cache block. It iterates through and calculates the minimum Euclidean distance between each trajectory vector in the current environment context vector and the historical expert demonstration trajectory matrix to establish the state trajectory envelope drift distance. This drift distance is then compared in real time with the boundary threshold in the system register. When extreme conditions cause the state trajectory envelope drift distance to exceed the preset boundary threshold, the system suspends the current forward computation flow and routes the state trajectory envelope drift distance as a nonlinear penalty scalar back to the multi-objective optimization layer. The multi-objective optimization layer increases the preset baseline mutation probability value of the non-dominated sorting genetic algorithm from 0.05 to the mutation probability threshold of 0.25 and proportionally amplifies the local adjustment step size of the feasible solution repair operator. This allows the multi-objective optimization layer to perform local iterative optimization within a compressed search window, centered on the current drifted controlled environment state. It then re-outputs a set of local expert trajectory fragments that fit the current abnormal conditions and inputs them into the control flow.

[0046] At this point, the central processing unit calls the loss function calculation unit to calculate the expert behavior clone gradient vector based on the local expert trajectory fragment, and simultaneously reads the current policy update gradient vector calculated based on the time-series differential error in the current control cycle. Subsequently, the system initiates a policy evolution constraint mechanism based on hyperspace gradient orthogonal projection. In the gradient clipping and projection register module of the dedicated calculation unit, the current policy update gradient vector array and the expert behavior clone gradient vector array, both with a dimension of 64, are multiplied element-wise and accumulated to obtain an inner product scalar value representing the degree of overlap between the two sets of vectors. Then, this inner product scalar value is divided by the square norm scalar obtained by accumulating the squares of each element in the expert behavior clone gradient vector array to calculate a dimensionless projection coefficient. This projection coefficient is then multiplied by each element of the expert behavior clone gradient vector array to generate a one-dimensional same-direction component array. Finally, each element in the current policy update gradient vector array is used to subtract the corresponding element in the one-dimensional same-direction component array. By completely eliminating collinear interference through this vector subtraction operation, the actual updated gradient vector is calculated and the policy network weight parameters are refreshed in situ. During the hyperspace update of the policy network, the geometric angle between the actual updated gradient vector obtained by subtraction and the expert behavior clone gradient vector is locked at a precise 90°, and the inner product of the two is always 0. Thus, while adapting to the time-delay control reorganization under unknown edge disturbances, it also prevents the overwriting of existing scheduling experience and catastrophic forgetting caused by heterogeneous data mixed updates.

[0047] Once the refresh is complete, the system automatically unsuspends the process and resumes forward propagation computation. A comprehensive evaluation function is set as the reward function to guide the optimization of the policy network weight parameters. The discrete expression formula for the reward function is as follows: Where R is the reward value, ΔF is the change in the overall task completion rate of the system after the rescheduling action, and its value ranges from -1 to 1, ΔS score For the predicted increment of satisfaction scores for affected customers, ΔT delay The system computational latency consumed by rescheduling decisions is considered. w1, w2, and w3 are globally locked constant coefficient scalars used to adjust the weighting of various business objectives. The agent obtains reward feedback through continuous trial and error with non-ideal environments. Finally, the forward propagation computes and outputs an adaptive real-time dynamic scheduling control matrix. The real-time dynamic scheduling control matrix is ​​established as a two-dimensional row-column structure, where the number of rows corresponds to the total number of tasks to be assigned M, the number of columns corresponds to the total number of personnel available for allocation N, and each intersection element in the matrix represents the joint scheduling matching probability value between the corresponding task and the corresponding personnel. Based on the joint scheduling matching probability value, the platform issues precise personnel dispatch control instructions, which shortens the dynamic rescheduling response processing time by more than 90% compared to manual intervention and increases the absolute value of the final task completion rate after rescheduling by more than 20%, enabling the system to maintain global convergence stability and continuous control flow operation when facing high-frequency extreme unknown disturbances.

[0048] Example 2: When a distributed intelligent computing service and a cross-regional personnel scheduling and management environment face the scheduling and orchestration of high-frequency burst task flows, the test platform uses a computing device with a main frequency of 3.2GHz and a memory capacity of 64GB. It reads a structured dataset consisting of 10,000 discrete task samples and 500 personnel profiles from the temporary storage area via the data bus. The data bus throughput is greater than or equal to 100MB / s, and the benchmark sampling frequency is not less than 50Hz. The system artificially superimposes burst electromagnetic and network environmental noise of a specific intensity into the benchmark test sequence. The system controls the random interference rate of sudden order cancellations by flexible workers to 15.4%, and injects data distortion items containing random typos and missing sentences into the text acquisition channel to simulate a controlled test state with severe covariate shift in the environmental context vector. The set value of the mutation probability of key control parameters in the memory of the computing device is used to balance the global trial-and-error exploration capability of the decision network in the wide-area state space with the convergence speed around the known path. When the input environmental context vector exhibits high-frequency oscillation characteristics due to sudden surges in demand, the mutation probability is graded between the preset basic mutation probability value of 0.05 and the mutation probability threshold of 0.25 to estimate the variance of the value function caused by random exploration in the forward propagation calculation of the control strategy network. When the calculated state trajectory envelope drift distance is greater than the preset boundary threshold of 1.20 in the system register, the mutation probability is adjusted to the extreme point of 0.25 so that the multi-objective optimization layer can complete the agile local evolution for the local distorted state and construct a closed deterministic parameter control loop.

[0049] To evaluate the bias convergence characteristics of the system under different step disturbances, the test system divided the controlled platform into control groups containing different combinations of features, with the first control group serving as a feature-deficient control group. While maintaining the same five-dimensional dynamic profile vector input, the expression formula of the five-dimensional dynamic profile vector is as follows: V=[S,C,A,P,G], where V is the five-dimensional structure feature vector, S is the skill index, C is the reputation index, A is the usability index, P is the preference index, and G is the growth index. The control group removes the policy evolution constraint mechanism based on hyperspace gradient orthogonal projection, making its parameter update dependent on the conventional temporal difference error gradient. The second control sample group serves as the parameter out-of-range control group, and its mutation probability threshold is set to a value of 0.50 exceeding the upper limit of the working window and a value of 0.01 below the lower limit. The sample group of this invention maintains the linkage state of feasible solution repair operator, state trajectory envelope self-calibration, and gradient space orthogonal constraint. The above sample groups are simultaneously deployed in low, medium, and high problem intensity scenarios with task demand transient surge rates of 10.5%, 35.2%, and 65.8%, respectively.

[0050] When the transient surge rate of task demands in high-intensity scenarios reaches 65.8%, the external task baseline shifts significantly. The state trajectory envelope drift distance of the first control sample group without a calibration loop rises to 3.14 within 5 consecutive sampling periods. Due to the lack of orthogonal projection constraints, the gradient vector of this sample group's online policy update contains overwrite components that contradict the direction of the gradient vector cloned from historical expert behavior, resulting in the overwriting of the network's existing weights. The value function estimation error changes by 42.5%. In contrast, the sample group of this invention establishes the state trajectory envelope drift distance as a scalar of 2.85 by traversing the processor before forward propagation calculation. Because the scalar value exceeds the preset boundary threshold of 1.20 in the system register, a suspension instruction is triggered. The system uses the state trajectory envelope drift distance as a penalty term and inputs it inversely into the multi-objective optimization layer. The mutation probability is switched from the base value of 0.05 to the upper limit of the working window of 0.25, and the adjustment step size of the feasible solution repair operator is proportionally increased, so that the multi-objective optimization layer converges within a 12.5ms time window and outputs a local expert trajectory fragment that fits the current offset state. The system immediately binds this to the supervisory control flow. At this time, the loss function calculation unit calculates the expert behavior clone gradient vector based on the local expert trajectory fragment, and synchronously reads the current policy update gradient vector calculated based on the temporal difference error. The current policy update gradient vector is projected onto the orthogonal complement space of the expert behavior clone gradient vector to eliminate the random exploration interference component caused by system noise and maintain the stability of the distribution of the updated weight parameters in the policy hyperspace.

[0051] During the final stage of the allocation iteration, the number of rows output by the system's data acquisition and computing devices corresponds to the total number of tasks to be allocated. And the number of columns corresponds to the total number of personnel that can be deployed. A two-dimensional real-time dynamic scheduling control matrix is ​​used, where each cross element in the adaptive matrix represents the joint scheduling matching probability value between the corresponding task and the corresponding personnel. The system thus records closed-loop performance characteristic data reflecting its operational status. Test data shows that for the second control group where the mutation probability threshold is set to an extreme value of 0.50, when the value exceeds the upper limit of the defined working window, the decision network's rate of improvement in task completion saturates. Furthermore, due to the disordered and discrete destruction of gene coding caused by the mutation operator, the system computation delay consumed by the rescheduling decision exhibits a non-linear and sharp increase. The formula for calculating the system computation delay is as follows: ΔT delay =T end -T start , where ΔT delay The system computation delay T consumed by the rescheduling decision end To control the output completion timestamp, T startThe environmental anomaly trigger timestamp increases the latency from 8.2ms under normal conditions to 38.6ms, delaying the system's steady-state adjustment response. When the mutation probability threshold is set to an extreme lower limit of 0.01, the multi-objective optimization layer lacks an in-situ local evolutionary adjustment trajectory when encountering a high-intensity environmental disturbance (65.8%). It cannot calculate the personnel replacement chromosome within the time limit for sudden order cancellations, causing the reward value obtained by the policy network in the forward feedback to degenerate into a negative value. The discrete expression formula for the reward value is as follows: Where R is the reward value, ΔF is the change in the overall task completion rate of the system, and ΔS is the value of the reward. score To predict the increment of customer satisfaction scores, w1, w2, and w3 are globally locked constant coefficient scalars. This negative reward triggers oscillations in the decision parameters, proving that the parameter boundary of 0.05 to 0.25 constitutes an optimization working window to maintain system convergence. In contrast, the sample group using the complete control method of this invention maintains an adaptive rescheduling response time of less than 12.0 ms when facing the same environmental disturbances. This performance reduces the time consumption by more than 90% compared to the traditional static scheduling method. Moreover, under the three gradient levels of transient surge rate of task demand at 10.5%, 35.2%, and 65.8%, the overall task completion rate after rescheduling remains at a high level of 98.2%, 96.4%, and 94.5%, respectively. The change in the overall task completion rate of the corresponding system, ΔF, always maintains a positive convergence trend, maintaining the continuous distribution of control flow in the scheduling control strategy network under heterogeneous noise interference.

[0052] Example 3: When a distributed intelligent computing service and a cross-regional personnel scheduling and management environment face a sudden surge of short-term, high-density task flows accompanied by high-frequency fluctuations in network channel latency, the computing device continuously reads concurrent scheduling transaction request information from the system data bus. At this time, due to unknown abrupt loss or grammatical tag corruption in the transmission network of unstructured text data sequences, the pre-trained large language model deployed in the contiguous memory address space experiences logical overflow when calculating its internal self-attention weight matrix. The forward computation tensor cannot output the corresponding structured dataset within the preset 5.0ms inherent clock cycle, causing a prolonged suspension and yielding of the platform's global control flow. The dispatching frequency of the controlled platform experiences an uncontrolled and sudden drop, leading to disordered branching and inconsistent configuration states in the computing service flow, resulting in technical problems of overall structural control flow obstruction and long-term computational losses.

[0053] To mitigate computational latency fluctuations and achieve closed-loop convergence of the computational state, the computing device initiates an algorithm-level abnormal branch flow control procedure. The specific operation process is as follows: First, the compliance detection unit monitors the computation time of the pre-trained large language model during forward mapping in real time. If the time exceeds the preset 5.0ms circuit breaker clock threshold, the compliance detection unit sends an interrupt wake-up signal to the central processing unit, triggering the circuit breaker strategy and suspending the forward computation flow of the pre-trained large language model. The central processing unit redirects the memory stack pointer. The original unstructured task requirement text stream is routed in real time to the input interface register of the parallel mounted Conditional Random Field (CRF) model. The CRF model uses a feature transition weight matrix pre-frozen in non-volatile memory and performs probabilistic statistical decoding of the text sequence labels using the Viterbi algorithm. Within a 1.5ms clock cycle, it directly produces a lower-level structured dataset containing the task type and physical spatial location coordinates and writes it back to the system data bus. This eliminates computational anomalies caused by text syntax corruption and completes the environmental scheduling requirement data.

[0054] Next, a dedicated computing unit reads the five-dimensional dynamic profile vectors of flexible workers from the lower-level structured dataset and cache block. A non-dominated sorting genetic algorithm for multi-objective optimization is then applied, where the skill and reputation indices in the personnel profile vectors are updated using the following adaptive dynamic calibration formula: , of which S new For the calibrated skill indicators, S old For the original skill index, α drift Let ΔT be a dimensionless stability decay coefficient used to characterize the negative inhibitory effect of scheduling delay on personnel allocation efficiency, set to 0.02. delay To mitigate the system computational latency consumed by rescheduling decisions, in the multi-objective optimization iterative evolution, facing the hard constraint of a daily cumulative working time limit of 8 hours, the processor traverses each gene position in the chromosome sequence in a time sequence. Once it detects that the cumulative working time of a specific agent within the current time window exceeds the 8-hour limit, the feasible solution repair operator is invoked to reverse-search for conflicting gene positions and forcibly replace their values ​​with available agent numbers, thereby outputting the optimal solution set that satisfies the physical constraints as a historical demonstration trajectory matrix.

[0055] To further confirm the compliance of the source of the preset boundary threshold value of 1.20 in the state trajectory envelope self-calibration, the dedicated computing unit ran a standard calibration program offline. This program imported multiple sets of historical evolution scenario samples with known physical residual gradients to calculate the impact trends of different drift distance settings on the overall system task completion rate variation and rescheduling decision computation latency. Data testing showed that when the boundary threshold was set below 1.20, the system overreacted to minor environmental noise, causing frequent suspensions of the multi-objective optimization layer and high-frequency triggering of local agile micro-evolutions, resulting in overload of the computing device's processor core and a sharp nonlinear increase in rescheduling computation latency to 45.0 ms. Conversely, when the boundary threshold was set above 1.20, the control policy network exhibited a hysteresis in perceiving already occurred covariate shifts. This causes the policy parameters to deviate in the update direction in the uncovered state space, resulting in a significant increase in the value function update variance. Ultimately, a value of 1.20 was determined as the basic boundary value for balancing computational delay loss and state envelope stability requirements. After completing the parameter calibration and circuit breaker branch flow control, the policy network performs hyperspace orthogonal projection on the generated actual update gradient vector. The system outputs an adaptive real-time dynamic scheduling control matrix containing a two-dimensional row and column structure through forward propagation. This allows the platform to issue precise personnel dispatch control commands based on the task and personnel joint scheduling matching probability values ​​in the matrix. The physical characteristic data recorded in the final test shows that, even under combined interference scenarios involving text syntax corruption and channel delay fluctuations, the system's rescheduling computation response delay is stably controlled within 12.0 ms. Furthermore, during long-term operation, the overall task completion rate variation remains within a positive and predictable range, enabling the distributed intelligent computing service platform to spontaneously exhibit a stable operating state characterized by dynamic conservation of computing resources and continuous distribution of control flow.

[0056] Example 4: When the system faces extreme scenarios with high-frequency expansion of computing service nodes and noise disturbances in the environmental context vector, the device runs a standardized pre-calibration procedure before the control flow to establish an initial state baseline. Specifically, the processor reads a standard feature mapping matrix from memory. The processor constrains the initial availability and skill indicators of personnel to numbers between 0 and 1 to construct a profile feature vector. The formula for expressing the profile feature vector is as follows: , where V base Using the baseline profile feature vector, S0 is the initial skill index, C0 is the initial reputation index, A0 is the initial usability index, P0 is the initial preference index, and G0 is the initial growth index. This provides a self-consistent input source for the online multi-objective optimization layer, avoiding data distribution drift caused by node deployment.

[0057] In specific incremental node access tests, the device injects the baseline profile feature vector into the policy network register. Before forward propagation, the processor calculates the minimum Euclidean distance between the current environment context vector and each trajectory vector in the historical expert demonstration trajectory matrix, and the resulting scalar value serves as the calibration boundary. The system completes convergence calculations offline. The results show that the state trajectory envelope drift distance established through the pre-calibration procedure is stably maintained within the normal range below the preset boundary threshold of 1.20. At this time, the polarity of the actual update gradient vector output by the policy network remains stable, enabling the real-time dynamic scheduling control matrix of the two-dimensional structure to output a deterministic joint scheduling matching probability value the moment the node comes online. This keeps the system computation latency consumed by rescheduling decisions within 12.0ms, and keeps the overall task completion rate of the system in a stable operating state with positive maximum convergence.

[0058] Example 5: In a distributed intelligent computing service and cross-regional personnel scheduling management scenario, the system faces overlapping tasks with high-frequency instantaneous fluctuations. The central processing unit (CPU) of the computing device reads unstructured task requirement text from the memory buffer. The CPU calls a pre-trained large language model to calculate the internal self-attention weight matrix of the input unstructured task requirement text and parses the implicit constraint information. The system data bus receives the task type, physical spatial coordinates, time window limits, skill qualification requirements, and basic compensation standards extracted by the large language model and converts them into a structured JSON dataset. Simultaneously, the system calls a compliance detection unit to determine the integrity of the structured JSON dataset.

[0059] Specifically, the compliance detection unit uses a Boolean mask matching mechanism to determine the integrity of structured data. The unit pre-stores a set of binary control words representing required fields. When the large language model outputs a structured dataset, the compliance detection unit extracts its key name set through the parser. This unit performs a bitwise logical AND operation on the extracted key name set and the binary control word. If the operation result shows that the required core key fields are missing, or the corresponding values ​​of the fields exceed the preset physical boundaries, resulting in data distortion, the parser performs real-time strong value range validation on the key business fields in the structured JSON dataset in the data boundary validation and format review register of the compliance detection unit: For the physical spatial location coordinate field, it checks whether its geographical longitude value is between 105.0°E and 125.0°E, and whether its latitude value is between 20.0°N and 40.0°N. If it exceeds this closed interval, it is established as spatial geometric distortion; For the time window boundary field, it calculates the difference between its end timestamp and start timestamp, and checks whether the task duration is between 30 minutes and 1440 minutes. If the difference is negative or exceeds 24 hours, it is established as time sequence distortion; For the basic compensation standard field, it checks whether its hourly unit price value is lower than 25.0 yuan or higher than 10000.0 yuan. If the result exceeds this physical reasonable boundary, it is established as an economic dimension distortion; once the verification result of any of the above fields touches the boundary, the status bit of the data verification register will monotonically transition from the normal status value 0 to the abnormal status value 1, thereby triggering the output of a clear data distortion flag bit.

[0060] The compliance detection unit sends a high-priority circuit breaker interrupt request to the central processing unit (CPU) via the hardware interrupt control bus. Upon receiving the interrupt signal, the CPU saves the process context of the current large language model during forward computation, forcibly suspends the currently occupied large language model's memory stack pointer, and stops transmitting computation instructions to the large language model. The CPU quickly resets the memory stack pointer to the starting address of the input interface register of the parallel-mounted Conditional Random Field (CRF) model. The CPU also activates the clock enable signal of the CRF model, thereby achieving time-sharing reuse and deterministic hard circuit breaker switching between heterogeneous models at the hardware and control protocol level. When a core critical field is missing or data distortion occurs in the structured JSON dataset, the compliance detection unit triggers the circuit breaker mechanism, shutting down the large language model and switching to the CRF model. The CRF model identifies sequence labels through time-series conditional probability statistics, produces a lower-level structured dataset containing task type and physical spatial location coordinates, and writes it back to the system data bus to complete the environmental scheduling requirement data.

[0061] When the compliance detection unit triggers the circuit breaker mechanism and shuts down the large language model to switch to the Conditional Random Field (CRF) model, the CRF model uses time-series conditional probability statistics to identify sequence labels and obtain a lower-level structured dataset containing task type and physical spatial location coordinates. For the three attribute fields missing from the structured JSON dataset due to model switching—time window limits, skill qualification requirements, and basic compensation standards—the system data bus automatically calls the historical sliding window matching module stored in non-volatile memory. This module uses the physical spatial location coordinates in the lower-level structured dataset as the center, retrieves historical scheduling records belonging to the same task type within a preset spatial radius of 5 kilometers, extracts the attribute averages of the historical scheduling records over the past 30 days as default configuration parameters to fill the structured JSON dataset, and delivers a complete and well-structured JSON dataset with comprehensive data features to the upstream multi-objective optimization layer and the downstream control strategy network, maintaining the computational continuity of the entire system's rescheduling control flow.

[0062] Before forward propagation computation, the central processing unit (CPU) reads environmental scheduling requirement data from the system data bus, extracts feature indicators including the number of tasks, personnel distribution, and collaborative boundary constraints, and calculates a high-dimensional feature vector based on these features through a fully connected layer, serving as the current environmental context vector. Specifically, before sending the feature indicators to the fully connected layer, the system initiates cross-dimensional dimension alignment and normalization processing. For the number of tasks, which is a scalar count, the system uses the historical maximum concurrent task threshold of 1000 for the current region as the baseline denominator and performs linear max-min normalization to map it to a dimensionless floating-point number between 0 and 1. For personnel distribution, which is a spatial geometric location, the system calculates the centroid distance loss of all available personnel agents relative to the center of the task core group using a discrete Euclidean two-dimensional geofence coordinate system, and uses a maximum service radius of 50 kilometers for boundary constraint truncation and scaling. For collaborative boundary constraints that are logical hard constraints, they are transformed into discrete feature vectors with deterministic Boolean values ​​of 0 or 1 through a pre-defined one-hot encoding. The system concatenates the normalized task quantity, personnel distribution scaling value, and collaborative boundary constraint feature vectors in a contiguous memory buffer to form a one-dimensional aligned multidimensional composite feature array. This array is then subjected to high-dimensional linear mapping through a fully connected layer with 64 neurons and layer normalization network operators. This ensures that the output high-dimensional feature vectors are mathematically and logically equivalent and compatible with the trajectory vectors in the historical expert demonstration trajectory matrix. The CPU retrieves the offline-generated expert optimization scheduling scheme sequence from the historical demonstration database and converts it into corresponding state-action trajectory vectors, dividing the data into multiple sub-storage blocks to form the historical expert demonstration trajectory matrix. The CPU iterates through and calculates the minimum Euclidean distance between the current environment context vector and each trajectory vector in the historical expert demonstration trajectory matrix. In the state-space topology analysis and distance retrieval register, the CPU retrieves the 128-dimensional current environment context vector and calculates the square of the difference between this vector and each 128-dimensional expert trajectory vector in the historical expert demonstration trajectory matrix across all dimensions. This difference is then accumulated across all dimensions and the square root operation is performed. This process iterates through and generates a set of discrete point-to-point geometric distance scalar values. Then, a minimum value retrieval operator is called to extract the absolute minimum value from this set. In mathematical topological mapping, this absolute minimum value is explicitly defined as the shortest geometric normal distance from the current dynamic operating condition state point to the multidimensional topological envelope surface of the historical stable expert control strategy. This seamlessly establishes a precise mapping from scalar distance to the degree of drift of the multidimensional envelope boundary, and directly outputs it as the established state trajectory envelope drift distance. In the first set of boundary test conditions, the minimum Euclidean distance between the input first environmental context vector and each trajectory vector in the historical expert demonstration trajectory matrix is ​​calculated to be 1.15. This value is less than the preset boundary threshold of 1.20, and the system directly uses depth... The network architecture drives the forward propagation computation of the control strategy network. In the second set of boundary test cases, the input second environment context vector undergoes covariate shift due to the influx of sudden task flows. The minimum Euclidean distance calculated by the analyzer in the central processing unit is 1.25. This value is greater than the preset boundary threshold of 1.20, and the central processing unit triggers a suspension instruction, suspending the current forward computation flow.

[0063] The central processing unit (CPU) uses the state trajectory envelope drift distance as a penalty to route back to the multi-objective optimization layer, updates the mutation probability threshold of the non-dominated sorting genetic algorithm, and proportionally amplifies the local adjustment step size of the feasible solution repair operator. In the discrete chromosome iterative repair control register of the multi-objective optimization layer, since the non-dominated sorting genetic algorithm uses discrete integer sequences composed of task node indices and corresponding personnel numbers to encode chromosomes, the local adjustment step size here is white-boxed and defined as the maximum discrete number of conflicting gene positions that can be simultaneously retrieved and forcibly replaced along the chromosome in a single repair control cycle. When the input state trajectory envelope drift distance is below 1.20, this local adjustment step size is assigned a normal baseline discrete value of 1. The system corrects and replaces only one gene locus exceeding the daily cumulative working hours limit at a time. However, when extreme disturbances cause the state trajectory envelope drift distance to jump to 2.85, the processor, through linear scaling, rigidly amplifies the local adjustment step size to the target discrete value of four. This drives the repair operator to simultaneously replace and correct the four gene loci violating the working hours inequality constraint along the chromosome encoding memory block within a single computation cycle. This generates a highly robust candidate solution set in a closed loop within the discrete solution space. The mutation probability threshold is updated by summing the preset baseline mutation probability value of 0.05 with the product of the adjustment coefficient and the state trajectory envelope drift distance. In the actual system debugging and offline calibration process, the aforementioned adjustment coefficient is fixed at a value of 1.60. The range of this adjustment coefficient is determined by introducing 20 sets of scene samples with known trajectory offsets for offline grid search.

[0064] When the adjustment coefficient is below the lower limit of 1.00, the system's response to environmental covariate shifts is too sluggish, causing the calculated mutation probability threshold to fail to cross the basic random exploration blind zone in time. This prevents the multi-objective optimization layer from reconstructing effective local expert trajectory segments within the preset 15.0ms clock cycle under sudden large disturbances. When the adjustment coefficient is above the upper limit of 2.50, even slight environmental channel fluctuations can cause an excessive jump in mutation probability, thereby disrupting the local convergence stability of the non-dominated sorting genetic algorithm around known good paths. This fluctuation caused a sharp increase in the nonlinear computational delay of the rescheduling decision to exceed 45.0 ms. By setting this coefficient to 1.60, it can be ensured that when the current state trajectory envelope drift distance is 0.125, the updated mutation probability precisely matches the upper limit of the working window of 0.25, achieving a dynamic balance of adaptive agile evolution. When determining the mutation probability threshold of the non-dominated sorting genetic algorithm in the multi-objective optimization layer, the processor uses a quantitative transformation path including a boundary truncation operator to calculate the final mutation probability. The mutation probability threshold is selected as the smaller value between the upper limit of the working window of 0.25 and the dynamically calculated value. The dynamically calculated value is generated by summing the preset baseline mutation probability value of 0.05 and the constant adjustment coefficient of 1.60 multiplied by the state trajectory envelope drift distance. When the input state trajectory envelope drift distance is 0.125, the dynamically calculated value obtained by summing the baseline mutation probability value of 0.05 and the product value of 0.20 is 0.25, and the output mutation probability threshold is selected as 0.25. When the system encounters severe covariate shift, causing the state trajectory envelope drift distance to rise to 2.85, although the dynamic calculated value obtained by summing the baseline mutation probability value of 0.05 and the product value of 4.56 reaches 4.61, the output mutation probability threshold remains fixed at the upper limit of the working window of 0.25 due to the hard truncation constraint of the boundary truncation operator. This reduces the parameter variance of the multi-objective optimization layer during local agile evolution. Under the second set of boundary test conditions, the calculated mutation probability threshold is adjusted from the baseline mutation probability value of 0.05 to the upper limit of the working window of 0.25. The multi-objective optimization layer uses integer sequence encoding to represent the gene chromosome of candidate solutions. The position index in the sequence corresponds to the specific task node, and the value at the position corresponds to the assigned human agent number.

[0065] During the crossover and mutation evolution process, the feasible solution repair operator modifies the solution space structure of infeasible solutions to address situations that violate supply-side spatial distribution constraints and heterogeneous individual attribute limitations. In the temporal time matrix verification and data filtering partitioning of the repair operator, the processor uses the allocation personnel number at each position in the current discrete integer encoded chromosome as the primary key to retrieve the real-time cumulative time statistics table stored in the system cache. The processor also performs a full discrete accumulation of the duration of all overlapping tasks assigned to that personnel within the current control cycle. Once it detects that the total cumulative time value of a personnel agent in a single day reaches and exceeds the preset hard limit of 8 hours of cumulative time in a single day, the repair operator immediately starts the reverse pointer retrieval program, performs a reverse traversal along the continuous memory address of the chromosome storage block to lock the conflicting gene position, and then retrieves the alternative available personnel list pre-stored in non-volatile memory with a service radius within 5km. The system filters out agent IDs that are idle within the current time window and whose cumulative working hours after replacement are still less than or equal to 8 hours. The central processing unit performs an in-situ memory overwrite operation, forcibly replacing the old ID at the original conflicting position with the selected new agent ID. Through this forced replacement and correction of discrete memory values, overload and spatial mismatch of working hours are eliminated, thereby outputting a discrete solution sequence that is completely legal in the data flow topology. When the cumulative working hours of a certain agent within the current time window after the mutation operation is 8.5 hours, which is greater than the preset daily cumulative working hour limit of 8 hours, the feasible solution repair operator searches for conflicting gene positions in reverse along the chromosome sequence and forcibly replaces its value with the agent ID with a cumulative working hour of 7.5 hours within the current time window and the second highest matching degree, so that the adjusted cumulative working hours are less than or equal to the daily cumulative working hour limit of 8 hours, producing a candidate solution set that satisfies the boundary hard constraints.

[0066] The multi-objective optimization layer uses the dynamic requirements extracted from the large language model as the current state center. It searches for the optimal solution with the smallest difference from the historical expert demonstration trajectory matrix within the candidate solution set, reconstructs local expert trajectory fragments, and writes them into a temporary storage matrix. The central processing unit (CPU) calls the loss function calculation unit to calculate the first derivative of the behavior cloning loss function with respect to the policy network weight parameters based on the local expert trajectory fragments, generating expert behavior cloning gradient vectors. Simultaneously, the CPU reads the current policy update gradient vector calculated based on the temporal difference error in the current control cycle. The CPU initiates a policy evolution constraint mechanism based on orthogonal projection of the hyperspace gradient, constructs the orthogonal complement space of the expert behavior cloning gradient vectors, projects the current policy update gradient vector onto the orthogonal complement space to obtain orthogonal gradient components, and calculates the actual update gradient vector. Specifically, the calculation formula for the actual update gradient vector is as follows: , where g act To update the gradient vector in high-dimensional reality; g tsUpdate the gradient vector for the current policy in higher dimensions; g bc For high-dimensional expert behavior cloning gradient vectors; This is the scalar product operator; `-` represents the square norm scalar of the expert behavior clone gradient vector; `-` represents the vector subtraction operator; ` / ` represents the division operator. When calculating the actual updated gradient vector, the computing device calls the loss function calculation unit to calculate the high-dimensional expert behavior clone gradient vector and the current policy update gradient vector. The system constructs an orthogonal complement space of the expert behavior clone gradient vector, projects the current policy update gradient vector onto the orthogonal complement space to obtain orthogonal gradient components. Specifically, it subtracts the component vector generated by dividing the scalar product of the current policy update gradient vector and the expert behavior clone gradient vector by the square norm of the expert behavior clone gradient vector and then multiplying it by the expert behavior clone gradient vector. The resulting vector difference is used as the in-situ refresh policy network weight parameters of the actual updated gradient vector. Since the inner product of the actual updated gradient vector and the expert behavior clone gradient vector is always equal to 0, the online updated gradient completely removes the adversarial components that contradict the direction of the existing expert experience gradient. This design allows the policy network to explore the solution space in situ in a hyperspace dimension orthogonal to the expert policy for edge unknown perturbations, and locks the existing weight distribution benchmark under standard conditions to prevent reverse overwriting. The central processing unit uses the calculated actual update gradient vector to perform in-situ high-frequency fine-tuning and refreshing of the weight parameters of the policy network, and restricts the online update direction to be in the same direction or orthogonal to the direction of the gradient vector cloned by the expert behavior.

[0067] The central processing unit (CPU) is unsuspended and resumes forward propagation computation. A comprehensive evaluation function is set as the reward function to guide the optimization of the policy network weight parameters. The formula for calculating the reward function is as follows: Where R is a dimensionless reward feedback signal scalar used to guide the updating of policy network weight parameters; ΔF is the change in the overall task completion rate of the system after the rescheduling action, a real number ranging from -1 to 1. score For the predicted increment of satisfaction scores for affected customers, ΔT delay The system computation delay consumed by the rescheduling decision is represented by w1, which is the first constant coefficient scalar of global locking; w2 is the second constant coefficient scalar of global locking; and w3 is the third constant coefficient scalar of global locking. + is the multiplication operator; + is the addition operator; - is the subtraction operator. In order to offset the magnitude gap between the three completely heterogeneous physical parameters—variable task completion rate, incremental customer satisfaction forecast, and system computation latency consumed by rescheduling decisions—this invention determines the first constant coefficient scalar to be 10.0, the second constant coefficient scalar to be 1.5, and the third constant coefficient scalar to be 0.2 through offline multi-criteria decision analysis calibration experiments.

[0068] Specifically, the first constant coefficient scalar is set to 10.0, which is used to amplify the variation in value between -1 and 1 to the order of tens, the same as the scoring item, thereby improving the sensitivity of the network strategy to the task completion rate constraint threshold; the second constant coefficient scalar is set to 1.5, which means that the customer satisfaction score, which is expressed as a score between 0 and 10, is linearly balanced and elastically adjusted; the third constant coefficient scalar is set to 0.2, which is used to reduce the computational delay of the system rescheduling decision, which is measured in milliseconds, to prevent the delay fluctuation of a few milliseconds to tens of milliseconds from generating excessive penalty, and to avoid the value function estimation error of the network strategy from exceeding 40% due to high-frequency spikes in the delay signal. This causes high-frequency oscillations in decision parameters. Through the specific numerical ratios mentioned above, cross-dimensional fusion and unification of heterogeneous feature data in a dimensionless space are achieved. The central processing unit applies the in-situ refreshed policy network weight parameters to the forward propagation calculation, outputting an adaptive real-time dynamic scheduling control matrix. The real-time dynamic scheduling control matrix is ​​configured as a two-dimensional row-column structure, where the number of rows corresponds to the total number of tasks to be assigned M, and the number of columns corresponds to the total number of personnel available for allocation N. Each intersection element in the matrix represents the joint scheduling matching probability value between the corresponding task and the corresponding personnel. The platform issues personnel dispatch control instructions based on the joint scheduling matching probability value to output dynamic scheduling control instructions. Under complex interference scenarios, the system's rescheduling calculation response latency remains stable within 12.0ms, and the overall task completion rate variation remains within the positive convergence range. The cross-regional personnel scheduling management platform exhibits a stable operating state with dynamic conservation of computing resources and continuous distribution of control flow.

[0069] The above description is merely a selection of preferred embodiments of this disclosure and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in the embodiments of this disclosure is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-described inventive concept. For example, technical solutions formed by substituting the above-described features with technical features having similar functions disclosed in the embodiments of this disclosure.

Claims

1. A method for intelligent scheduling and control in complex environments based on imitation learning, characterized in that, include: Step S101: Call the large language model to calculate the internal self-attention weight matrix of the input unstructured task requirement text, parse the implicit constraint information and extract the task attribute fields, convert it into a structured dataset and output it to the system data bus, synchronously call the compliance detection unit to determine the integrity of the structured dataset, and trigger the circuit breaker mechanism when encountering data distortion or failure of the large language model, and switch to the conditional random field model to obtain environmental scheduling requirement data. Step S102: Before forward propagation calculation, the processor reads the current environment context vector and retrieves the historical expert demonstration trajectory matrix. It then iterates through and calculates the minimum Euclidean distance between each trajectory vector in the environment context vector and the historical expert demonstration trajectory matrix to establish the state trajectory envelope drift distance. Step S103: When the state trajectory envelope drift distance is greater than the preset boundary threshold, suspend the current forward computation flow, route the state trajectory envelope drift distance as a penalty to the multi-objective optimization layer, increase the mutation probability threshold of the non-dominated sorting genetic algorithm and proportionally increase the local adjustment step size of the feasible solution repair operator, reconstruct the local expert trajectory segment with the current state as the center and refresh the network weights in situ. Step S104: Calculate the expert behavior clone gradient vector based on the local expert trajectory segment, synchronously read the current policy update gradient vector corresponding to the time-series difference error, project the current policy update gradient vector onto the orthogonal complement space of the expert behavior clone gradient vector, calculate and generate the actual update gradient vector, and refresh the policy network weight parameters in situ to output dynamic scheduling control instructions.

2. The intelligent scheduling and control method based on imitation learning in a complex environment according to claim 1, characterized in that, The task attribute fields in step S101 include task type, physical spatial location coordinates, time window limits, skill qualification requirements, and basic remuneration standards; the structured dataset is a structured JSON dataset.

3. The intelligent scheduling and control method based on imitation learning in a complex environment according to claim 2, characterized in that, Step S101 includes the following sub-steps: Step S1011: Use a large language model to identify the relationship between entity labels and physical constraints in unstructured task requirement text, and map them to generate a structured JSON dataset containing task requirement dimensions, spatial distribution coordinates, and individual attributes. In step S1012, when the integrity of the structured JSON dataset is determined to be incomplete, the compliance detection unit triggers the circuit breaker mechanism, shuts down the large language model and switches to the conditional random field model, and identifies sequence labels through time series conditional probability statistics to complete the environmental scheduling requirement data.

4. The intelligent scheduling and control method for complex environments based on imitation learning according to claim 1, characterized in that, Step S102 includes the following sub-steps: Step S1021: Read the current environment scheduling requirement data from the system data bus and extract feature indicators including the number of tasks, personnel distribution and collaborative boundary constraints. Map the feature indicators into high-dimensional feature vectors through a fully connected layer to serve as the current environment context vector. Step S1022: Retrieve the offline-generated expert optimization scheduling scheme sequence from the historical demonstration database and convert the expert optimization scheduling scheme sequence into the corresponding state action trajectory vector, and perform sub-storage block partitioning to form a historical expert demonstration trajectory matrix containing multiple sets of time-series associated states.

5. The intelligent scheduling and control method based on imitation learning in a complex environment according to claim 1, characterized in that, In step S103, the mutation probability threshold of the non-dominated sorting genetic algorithm is increased. Specifically, the preset baseline mutation probability value is summed with the product of the adjustment coefficient and the state trajectory envelope drift distance to update the mutation probability threshold of the non-dominated sorting genetic algorithm.

6. The intelligent scheduling and control method based on imitation learning in a complex environment according to claim 1, characterized in that, Step S103 includes the following sub-steps: Step S1031: Increase the local adjustment step size of the feasible solution repair operator, and use the feasible solution repair operator to map and correct the infeasible solution space that violates the spatial distribution constraints of the supply side and the attribute restrictions of heterogeneous individuals, so as to generate a candidate solution set that satisfies the hard boundary constraints; Step S1032: The dynamic requirements extracted by the large language model are used as the current state center. The multi-objective optimization algorithm is used to search for the optimal solution with the smallest difference from the historical expert demonstration trajectory matrix in the candidate solution set. The local expert trajectory segment is reconstructed and the network weights are refreshed in situ.

7. The intelligent scheduling and control method based on imitation learning in a complex environment according to claim 1, characterized in that, Step S104 includes the following sub-steps: Step S1041: Construct a behavior cloning loss function containing the first derivative term with respect to the policy network weight parameters, and calculate the first derivative of the behavior cloning loss function with respect to the policy network weight parameters based on the local expert trajectory fragments to generate the expert behavior cloning gradient vector. Step S1042: Synchronously read the timing difference error in the current control cycle, calculate the gradient of the timing difference error with respect to the value function to generate the current policy update gradient vector; Step S1043: Construct the orthogonal complement space of the expert behavior clone gradient vector, project the current policy update gradient vector onto the orthogonal complement space of the expert behavior clone gradient vector to obtain orthogonal gradient components, use the orthogonal gradient components as the actual update gradient vector and refresh the policy network weight parameters in situ, and restrict the online update direction to be in the same direction as the expert behavior clone gradient vector.

8. The intelligent scheduling and control method based on imitation learning in a complex environment according to claim 1, characterized in that, The process of updating the policy network weight parameters also includes the following steps: Step S105: Establish an experience replay buffer to store the status, actions, reward feedback signals and subsequent status data during the historical scheduling process; Step S106: Randomly extract a small batch of sample data from the experience replay buffer to decouple the temporal correlation of historical data; Step S107: Simultaneously construct the target value network and use the target value network to calculate the target value function value in order to maintain the stability of the value function update.

9. The intelligent scheduling and control method based on imitation learning in a complex environment according to claim 8, characterized in that, The reward feedback signal in step S105 is generated through the following sub-steps: Step S1051: Monitor the task completion quality of supply-side personnel and the service satisfaction of demand-side personnel in real time, and calculate the positive revenue reward value. Step S1052: Calculate the geographic distance loss and individual collaboration cost of heterogeneous personnel in the scheduling process, and generate a negative penalty term; Step S1053: The positive reward value and the negative penalty term are weighted and summed to output a reward feedback signal used to guide the update of the policy network weight parameters.

10. The intelligent scheduling and control method based on imitation learning in a complex environment according to claim 1, characterized in that, Step S104, which outputs the dynamic scheduling control instruction, includes the following sub-steps: Step S1044: Apply the in-situ refreshed policy network weight parameters to the forward propagation calculation and output an adaptive real-time dynamic scheduling control matrix. Step S1045: Establish a two-dimensional row and column structure of the real-time dynamic scheduling control matrix, where the number of rows corresponds to the total number of tasks to be assigned and the number of columns corresponds to the total number of personnel that can be deployed. Each cross element in the real-time dynamic scheduling control matrix represents the joint scheduling matching probability value between the corresponding task and the corresponding personnel. Based on the joint scheduling matching probability value, issue personnel dispatch control instructions to output dynamic scheduling control instructions.

Images