Intelligent Agent task scheduling and cultivation management system based on time sequence modeling

Abstract

The invention discloses an intelligent Agent task scheduling and cultivation management system based on time sequence modeling, and the system comprises a data collection and preprocessing module which is used for collecting task, resource and capability data and initializing a solution vector; the time sequence feasible region homotopy module is used for generating a feasible region and establishing a stage index; the constraint projection drawing module is used for constructing a constraint projection drawing and determining a leading edge set; the seed selection and direction calculation module is used for selecting a seed constraint and calculating a projection direction; the improved Kaczmarz module is used for updating and generating candidate solution vectors along the direction; the default integral evaluation module is used for calculating a time window integral and generating an acceptance mark; the state updating module is used for writing a solution vector and updating leading edge and stage indexes; and the plan generation module is used for outputting task assignment and resource allocation. According to the invention, the stability of scheduling solution and the time sequence control precision are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent scheduling, and particularly relates to an intelligent Agent task scheduling and management system based on time sequence modeling. BACKGROUND

[0002] Existing intelligent scheduling mostly adopts rule driving or mixed integer programming solving, models task time window, resource capacity and priority as static constraints, and generates assignment scheme in offline scene. When task arrival and execution time have obvious time sequence and uncertainty, common practice is periodic recalculation or local adjustment using heuristics, which is difficult to maintain stability between calculation overhead and scheme feasibility. For complex processes, the industry also introduces process dependency modeling and rolling prediction based on DAG, but the dependency relationship mostly exists as a scheduling precondition, lacking a tight coupling mechanism with the online solving kernel, resulting in a disconnection between constraint activation state and row selection strategy. In addition, numerical methods for high-dimensional sparse constraints are mostly gradient descent, coordinate descent or traditional Kaczmarz row projection, and the row selection is usually based on fixed order, equal probability or residual size, which is difficult to reflect the dynamic changes of time dependence and risk exposure.

[0003] To improve convergence and feasibility, some schemes attempt block projection, momentum acceleration, preconditioning or parallel processing, but these numerical enhancements are mostly developed around geometric or spectral properties, without mapping time sequence constraints, causal relationships and budget constraints into selection and acceptance criteria. Violation metrics are usually reflected in penalty terms or soft constraints, lacking explicit integration accounts and budget comparison mechanisms for time windows, resulting in uncontrollable long-term feasibility while short-term residual decreases. On the other hand, existing systems often serialize "prediction-scheduling-execution" as a loosely coupled process, with prediction only used for initialization or parameter setting, without forming a stage-by-stage contraction strategy based on stage index and homotopy path, resulting in solution trajectory oscillation, front advance stagnation or repeated rollback when load suddenly changes, affecting online availability and calculation certainty.

[0004] Therefore, how to provide an intelligent Agent task scheduling and management system based on time sequence modeling is a problem that those skilled in the art need to solve. SUMMARY

[0005] One purpose of the present application is to propose a time sequence modeling based intelligent Agent task scheduling and development management system, which fully utilizes the windowed time sequence feasible region homotopy, constraint projection graph and frontier set based on time dependence and causal relationship, seed selection and composite projection direction calculation, improved Kaczmarz projection update, and integrated mechanism of default integral-budget comparison and acceptance writing according to time window measurement, and constructs a time-varying feasible region contraction and online iterative solution process oriented to task arrival sequence, which can realize adaptive step and frontier advance collaborative control of solution update in stage dimension, so as to complete task assignment, start and end time determination, resource allocation and ability development arrangement, and has the advantages of high time sequence constraint modeling precision, clear online feasibility control, stable convergence under high-dimensional sparse constraint, and fast response to load fluctuation.

[0006] The time sequence modeling based intelligent Agent task scheduling and development management system according to the embodiment of the present application comprises: A data acquisition and preprocessing module is configured to acquire task data, resource data and Agent capability profile, perform time alignment and standardization, and initialize a current solution vector and a time sequence constraint set. A time sequence feasible region homotopy module is configured to generate a windowed time sequence feasible region based on the time sequence constraint set, construct a homotopy from an initial feasible region to a target feasible region and establish a stage index. A constraint projection graph module is configured to construct a constraint projection graph and determine a frontier set according to time dependence and causal relationship. A seed selection and composite projection direction module is configured to select a seed constraint from the frontier set within a stage, aggregate adjacent constraints according to adjacency relationship and calculate a composite projection direction. An improved Kaczmarz iteration module is configured to perform improved Kaczmarz projection update along the composite projection direction based on the current solution vector and generate a candidate solution vector. A default integral and budget evaluation module is configured to establish a default integral account book measured according to a time window, calculate a time window integral corresponding to the candidate solution vector and compare it with a budget, generate an acceptance mark, perform update scaling and seed replacement and recalculate the composite projection direction when the acceptance mark is not acceptable. An acceptance writing and state updating module is configured to write the candidate solution vector as the current solution vector when the acceptance mark is acceptable, update the frontier set, the stage index and the default integral account book, and register a reason and keep the current solution vector unchanged when the acceptance mark is not acceptable. A plan generation module is configured to generate task assignment, start and end time, resource allocation and ability development arrangement based on the current solution vector, and output scheduling and development plans.

[0007] Optionally, the modules are realized through the following methods: Collecting task data, resource data and agent capability profile, completing time alignment and standardization, initializing current solution vector and time sequence constraint set; Generating windowed time sequence feasible region from time sequence constraint set, constructing homotopy from initial feasible region to target feasible region and establishing stage index; Building constraint projection graph and determining frontier set based on time dependence and causality; Selecting seed constraint from frontier set at current stage, aggregating adjacent constraints according to adjacency relationship, and calculating composite projection direction; Performing improved Kaczmarz projection update along composite projection direction with current solution vector to generate candidate solution vector; Establishing a default integral ledger measured by time window, calculating the time window integral corresponding to the candidate solution vector and comparing it with the budget, marking it as an acceptable update when the time window integral does not exceed the budget, scaling the update amplitude when it does not meet the requirement, and replacing the seed constraint and recalculating the composite projection direction when it still does not meet the requirement; Adopting the acceptable update as the current solution vector, advancing the constraint projection graph frontier and updating the stage index and default integral ledger, and retaining the current solution vector for the unacceptable update and updating the default integral ledger; Generating task assignment, start and end time, resource allocation and capability development arrangement with the current solution vector, outputting the schedule and development plan when the termination condition is met, and continuing iteration on the updated stage index and frontier set when it is not met.

[0008] Optionally, the improved Kaczmarz includes: Composite projection direction generation, calculating weights from seed constraints and adjacent constraints and linearly combining row vectors to form a composite projection direction based on the constraint projection graph; Homotopy stage step control, mapping stage index to step upper limit and direction weight adjustment coefficient and loading into iteration; Half-space projection correction, performing half-space projection on inequality constraints and truncating out-of-bounds updates; Budget acceptance criterion, calculating the updated time window integral from the default integral ledger and comparing it with the budget, scaling the update amplitude when it exceeds the budget, replacing the seed constraint and recalculating when it still exceeds the budget, and accepting the update when it does not exceed the budget; Frontier advancement and stage update, advancing the constraint projection graph frontier and synchronously updating the stage index and default integral ledger when the update is accepted; Solution vector writing, recording the accepted candidate solution vector as the current solution vector and as the starting point for the next iteration.

[0009] Optionally, the construction of the time sequence feasible region and homotopy includes: Determine the time range and window length, divide it into consecutive windows in order and number them, and establish a window index set; assigning the time sequence constraint set to corresponding windows according to window index to form window-level constraint set and recording time window label, priority label and causality label; constructing initial feasible region based on window-level constraint set of initial window and recording boundary description and variable range; constructing target feasible region based on window-level constraint set of target window set and recording boundary description and variable range; setting homotopy parameter value sequence which monotonically increases from zero to one, determining stage number and establishing stage index; setting activation state and weight of window-level constraint set according to homotopy parameter in each stage, generating corresponding windowed time sequence feasible region and synchronously recording stage and window correspondence.

[0010] Optionally, the constraint projection graph construction and frontier set determination specifically includes: determining time precedence relationship and causality relationship between constraints according to window index set and time sequence constraint set to form relationship determination record; assigning window index k and row index i to each constraint, setting node label as“k-i”and establishing node list; formulating directed connection rule according to relationship determination record, limiting to establish directed connection from predecessor to successor when predecessor time is not later than successor and causality trigger exists; generating directed edge set according to directed connection rule to form constraint projection graph together with node list; generating adjacency relationship table according to node number to respectively register predecessor list and successor list of each node; setting activation label of each node in constraint projection graph according to stage index and homotopy parameter, limiting value to“activated”and“unactivated”; forming activated subgraph on node and corresponding directed edge with activated label as“activated”, counting in-degree of each node in activated subgraph as in-degree count; determining node set with in-degree count of zero and activated label of“activated”as frontier set and saving frontier set and corresponding adjacency relationship table in ascending order of node number.

[0011] Optionally, the seed selection and composite projection direction calculation specifically includes: selecting seed constraint from frontier set in current stage according to priority label, window index and node number in order, recording corresponding node number and window index; collecting first-order adjacent constraints according to directed connection relationship of constraint projection graph, first-order range containing direct predecessor and direct successor, intercepting adjacent constraint set according to preset upper limit and recording node number and window index of each constraint in set; An aggregate weight is calculated for each constraint in the adjacency constraint set, and the aggregate weight is composed of three components: a priority component, a stage component, and an adjacency distance component; the three components are multiplied to obtain an original weight, and normalization is performed on all original weights to make the weights non-negative and the sum of the weights equal to one; A constraint row vector corresponding to the adjacency constraint is extracted in ascending order of node number, the row vector has the same dimension as the current solution vector, and a mapping table of the row vector sequence and the node number is established; The row vector sequence is linearly combined with the normalized aggregate weight to obtain a composite projection direction, and an index mapping of the composite projection direction and the seed constraint is established; An amplitude clipping is performed to limit the length of the composite projection direction within a predetermined upper limit; A consistency check is performed to compare the composite projection direction with the previous composite projection direction in terms of directionality, and if the included angle is not greater than a threshold value, it is considered to pass; if the comparison fails, the aggregate weight is re-normalized and linearly combined again; if it still fails, the seed constraint is replaced and the adjacency collection, weight calculation and linear combination are repeated; After the consistency check passes, the composite projection direction is determined, submitted to the improved Kaczmarz projection update, and the node number, window index, aggregate weight and direction length are recorded for tracing within the stage.

[0012] Optionally, the execution of the improved Kaczmarz projection update specifically includes: The current solution vector, the composite projection direction, the node number sequence of the adjacency constraint set, the aggregate weight sequence, the constraint row vector sequence and the constant term sequence are read, and the dimension consistency check and the missing value cleaning are completed; The row residual is calculated for each constraint in the node number sequence, and the row residual is defined as the constant term minus the inner product of the constraint row vector and the current solution vector, and the row residual sequence is formed and the index correspondence relationship is recorded; The row residual sequence is weighted and summed using the aggregate weight sequence to obtain a weighted residual scalar, and the weighting order is consistent with the node number sequence; A curvature scalar in the composite projection direction is calculated, and the curvature scalar is defined as: the square of the projection length of each constraint row vector in the composite projection direction is weighted and summed according to the aggregate weight, and a positive stability term is added to the summation result; A step parameter is determined, and the step parameter is the quotient of the weighted residual scalar and the curvature scalar, and the denominator includes the stability term; if the step parameter is less than zero, it is set to zero, and if it is greater than a predetermined upper limit, it is set to the predetermined upper limit; The current solution vector is updated along the composite projection direction according to the step parameter to generate a candidate solution vector, and the numerical precision and the storage order are consistent with the current solution vector; Performing direction consistency check, consistency check defined as: the angle between the increment vector obtained by subtracting the current solution vector from the candidate solution vector and the composite projection direction is not greater than the threshold; If not satisfied, reduce the step size parameter by a certain proportion and update again until the consistency check is satisfied; If the step size parameter is reduced to the lower limit and still not satisfied, stop updating and keep the current solution vector unchanged; Performing amplitude clipping, amplitude clipping defined as: the Euclidean distance between the candidate solution vector and the current solution vector does not exceed the preset radius; If exceeded, scale the increment vector to make the distance equal to the preset radius; Recording the candidate solution vector, step size parameter, curvature scalar, weighted residual scalar, composite projection direction length, node number sequence, aggregation weight sequence, timestamp and stage index as the reading object of the default integral ledger calculation.

[0013] Optionally, the default integral ledger and budget comparison process specifically includes: Establishing a default integral ledger measured by time window, the ledger field containing time window index, budget value, current integral value, cumulative integral value, timestamp; Mapping the candidate solution vector with the constraint set within the time window, determining the constraint row vector and constant term participating in the calculation according to the node number sequence and aggregation weight sequence; Calculating row residual piece by piece, row residual defined as the inner product of the constant term minus the constraint row vector and the candidate solution vector, obtaining a row residual sequence; Taking the positive part of the row residual sequence, the positive part defined as zero for negative values, and weighted summing according to the aggregation weight sequence in each time window to obtain the time window integral increment; Updating the current integral value and cumulative integral value according to the time window, and recording the snapshot of the node number sequence, aggregation weight sequence, stage index and timestamp; Comparing the current integral value with the budget value piece by piece in each time window to generate two types of results, acceptance mark and non-acceptance mark; When there is a non-acceptance mark, calculating the scaling coefficient interval, selecting the minimum scaling coefficient that satisfies all time window budget constraints, and uniformly scaling the increment of the candidate solution vector and the current solution vector; After completing the uniform scaling, recalculating the time window integral increment and comparison result, if there is still a non-acceptance mark, performing seed constraint replacement and recalculating the composite projection direction according to the constraint projection graph to generate a new candidate solution vector, and then performing mapping, row residual calculation, time window integral increment calculation and comparison again; When all time windows are marked as accepted, an acceptable update is formed, the scaling coefficient, time window level integral, stage index, timestamp and related index information are registered, and the candidate solution vector is marked as an update that can be adopted.

[0014] Optionally, the acceptance writing and state updating specifically includes: read the acceptance mark, and simultaneously read the record of the candidate solution vector, the stage index, the timestamp, the scaling factor, the node number sequence and the aggregation weight sequence; when the acceptance mark is acceptable, record the candidate solution vector as the current solution vector, and register the stage index and the timestamp; in the constraint projection graph and the active subgraph, mark the seed constraint and the first-order adjacent constraint participating in the composite projection as processed, and remove the processed node from the frontier set; in the active subgraph, update the in-degree count, add the node with the in-degree count of zero to the frontier set, and keep the node number in ascending order; update the stage index according to the homotopy parameter sequence, and when there is no unprocessed node in the active subgraph, increment the stage index, and when there is an unprocessed node in the active subgraph, keep the stage index unchanged; write the default score ledger, which contains the time window score, the cumulative score, the stage index, the timestamp, the scaling factor, the node number sequence and the aggregation weight sequence; when the acceptance mark is not acceptable, keep the current solution vector unchanged, keep the frontier set unchanged, register the reason for non-acceptance, the associated time window index and the timestamp; generate a stage state record, which contains the stage index, the frontier set, the number of nodes in the active subgraph, the acceptance mark and the identification of the current solution vector, and after the record is completed, enter the next iteration.

[0015] The beneficial effects of the present application are: The present application cooperates the windowed time sequence feasible homotopy and the constraint projection graph, directly maps the time dependence and the causal relationship into the row-level selection and the direction construction, combines the seed selection driven by the frontier set and the composite projection direction, makes the solution update sequentially advance along the stage index, avoids the solution trajectory shock and the repeated rollback when the static or loosely coupled scheduling appears under the load fluctuation, improves the projection update of the Kaczmarz in the composite projection direction, controls the stepping strength by the stage index, and makes the online solution maintain the stable iteration path and the controllable computing overhead under the high-dimensional sparse constraint.

[0016] A unified acceptance rule is established based on the comparison between the default score ledger measured by the time window and the budget, the adoption of the candidate solution vector does not depend on the heuristic threshold, but is directly determined by the relationship between the time window score and the budget, and is linked with the acceptance writing and the state update, forming a closed-loop process from data collection and preprocessing to plan generation, compared with the existing scheme which only depends on periodic recalculation or local heuristic, the present application can output the adoptable task assignment, start and end time and resource allocation in the continuously arriving task stream, while taking into account the ability training arrangement, so that the time constraint modeling accuracy, online feasibility control and convergence stability are consistently improved. BRIEF DESCRIPTION OF DRAWINGS

[0017] The accompanying drawings are included to provide a further understanding of the application and are incorporated in and constitute a part of this specification, illustrate embodiments of the application and together with the description serve to explain the principles of the application. In the drawings:

[0018] Fig. 1 Flow chart of the intelligent Agent task scheduling and cultivation management system based on time sequence modeling proposed by the application; Fig. 2 System module block diagram of the intelligent Agent task scheduling and cultivation management system based on time sequence modeling proposed by the application; Fig. 3 Homotopy diagram of the time sequence feasible region of the intelligent Agent task scheduling and cultivation management system based on time sequence modeling proposed by the application. DETAILED DESCRIPTION

[0019] The application will now be described in further detail with reference to the drawings. These drawings show only the essential features of the application and are therefore to be regarded only as a schematic illustration. In the drawings:

[0020] REFERENCE Figs. 1-3 The intelligent Agent task scheduling and cultivation management system based on time sequence modeling comprises: A data acquisition and preprocessing module for acquiring task data, resource data and Agent capability profile, performing time alignment and standardization, initializing a current solution vector and a time sequence constraint set; A time sequence feasible region homotopy module for generating a windowed time sequence feasible region based on the time sequence constraint set, constructing a homotopy from an initial feasible region to a target feasible region and establishing a stage index; A constraint projection graph module for constructing a constraint projection graph and determining a frontier set according to time dependence and causal relationship; A seed selection and composite projection direction module for selecting a seed constraint from the frontier set within a stage, aggregating adjacent constraints according to adjacency relationship and calculating a composite projection direction; An improved Kaczmarz iteration module for performing improved Kaczmarz projection update along the composite projection direction based on the current solution vector and generating a candidate solution vector; A breach integral and budget evaluation module for establishing a breach integral account book measured by time window, calculating the time window integral corresponding to the candidate solution vector and comparing it with the budget, generating an acceptance flag, performing update scaling and seed replacement and recalculating the composite projection direction when the acceptance flag is not accepted; An acceptance writing and state updating module for writing the candidate solution vector as the current solution vector when the acceptance flag is acceptable, updating the frontier set, the stage index and the breach integral account book, and registering the reason when the acceptance flag is not acceptable and keeping the current solution vector unchanged; A plan generation module is configured to generate task assignments, start and end times, resource allocations and capability development schedules based on the current solution vector, and output a schedule and development plan.

[0021] The application realizes unified modeling and dynamic solving of task data, resource data and agent capability portrait by constructing an intelligent agent task scheduling and development management system based on time sequence modeling, adopts a collaborative mechanism of time sequence feasible region homotopy, constraint projection graph and improved Kaczmarz iteration, enables task scheduling and agent capability growth to be optimized synchronously in a multi-constraint time sequence environment, effectively improves the coordination, constraint satisfaction rate and resource utilization efficiency of task allocation, and realizes continuous feasibility and dynamic adaptive scheduling of the task execution process.

[0022] In the embodiment, the modules are realized by the following method: Collect task data, resource data and agent capability portrait, complete time alignment and standardization, initialize the current solution vector and the time sequence constraint set; Generate a windowed time sequence feasible region from the time sequence constraint set, construct a homotopy from the initial feasible region to the target feasible region, and establish a stage index; Based on the time dependence and causal relationship, construct a constraint projection graph and determine a frontier set; Select a seed constraint from the frontier set at the current stage, aggregate adjacent constraints according to adjacency relationship, and calculate a composite projection direction; Perform improved Kaczmarz projection update along the composite projection direction with the current solution vector to generate a candidate solution vector; Establish a default violation integral account measured by time window, calculate the time window integral corresponding to the candidate solution vector and compare it with the budget, mark it as an acceptable update when the time window integral does not exceed the budget, scale the update amplitude when it does not meet the requirement, and replace the seed constraint and recalculate the composite projection direction when it still does not meet the requirement; Adopt the acceptable update as the current solution vector, advance the constraint projection graph frontier and update the stage index and the violation integral account, and keep the current solution vector for the unacceptable update and update the violation integral account; Generate task assignments, start and end times, resource allocations and capability development schedules based on the current solution vector, output a schedule and development plan when the termination condition is met, and continue iteration on the updated stage index and frontier set when the termination condition is not met.

[0023] The application integrates task collection, constraint generation, topology construction, projection iteration, budget evaluation and result generation into a closed-loop control process through a modular time sequence solving process, so that the system can maintain calculation stability and convergence characteristics under complex time sequence constraints and multi-dimensional resource conditions, thereby ensuring the time sequence consistency and global optimality of the task scheduling and capability development process, and maintaining solving convergence and constraint feasibility in high-load dynamic task scenarios.

[0024] In this embodiment, the improved Kaczmarz includes: Composite projection direction generation, calculating weights and linearly combining row vectors to form a composite projection direction based on the constraint projection graph from the seed constraint and the adjacent constraint; Homotopy phase step control, mapping the phase index to the step length upper limit and the direction weight adjustment coefficient and loading it into the iteration; Half-space projection correction, performing half-space projection on the inequality constraint and truncating the out-of-bound update; Budget acceptance criterion, calculating the updated time window integral according to the default integral account and comparing it with the budget, scaling the update amplitude when it exceeds the budget, replacing the seed constraint and recalculating when it still exceeds the budget, and accepting the update when it does not exceed the budget; Frontier advance and phase update, advancing the constraint projection graph frontier when the update is accepted and synchronously updating the phase index and the default integral account; Solution vector writing, recording the accepted candidate solution vector as the current solution vector and as the starting point for the next iteration.

[0025] The improved Kaczmarz iteration method of the application introduces mechanisms such as composite projection direction generation, half-space projection correction and budget acceptance criterion, so that in a high-dimensional constraint system, the solution vector can be gradually approximated and the constraint can be converged with relatively low computational complexity. Combined with phase step control, the direction step length is adaptively adjusted, thereby improving the stability and accuracy of the scheduling solution in the scenario where time window constraints and resource budget constraints coexist.

[0026] In this embodiment, the construction of the time sequence feasible region and the homotopy includes: Determine the time range and window length, divide them into continuous windows in order and number them, and establish a window index set; Assign the time sequence constraint set to the corresponding window according to the window index, form the window-level constraint set, and record the time window label, priority label and causal label; Take the window-level constraint set of the starting window as the basis to construct the initial feasible region, and record the boundary description and variable range; Take the window-level constraint set of the target window set as the basis to construct the target feasible region, and record the boundary description and variable range; Set the homotopy parameter value sequence to monotonically increase from zero to one, determine the number of stages and establish the stage index; In each stage, set the activation state and weight of the window-level constraint set according to the homotopy parameter, generate the corresponding windowed time sequence feasible region, and synchronously record the correspondence between the stage and the window.

[0027] The application can realize stage solving control and smooth transition through homotopy parameters, ensure the continuous feasibility of the task time sequence constraints in the solving process, and significantly reduce the solving instability problem caused by constraint jump.

[0028] In the embodiment, the construction of the constraint projection graph and the determination of the frontier set specifically include: According to the window index set and the time sequence constraint set, the time sequence relationship and the causal relationship between the constraints are determined one by one to form a relationship determination record; A window index k and a row index i are assigned to each constraint, a node identifier is set as "k-i", and a node list is established; According to the relationship determination record, a directed connection rule is formulated, and when the predecessor time is not later than the successor and there is a causal trigger, a directed connection from the predecessor to the successor is established; According to the directed connection rule, a directed edge set is generated, and the constraint projection graph is formed together with the node list; An adjacency relationship table is generated according to the node number, and the predecessor list and the successor list of each node are respectively registered; In combination with the stage index and the homotopy parameter, an activation mark is set for each node in the constraint projection graph, and the value is limited to "activated" and "unactivated"; An activation subgraph is formed on the node with the activation mark "activated" and the corresponding directed edge, and the number of predecessors of each node in the activation subgraph is counted as the in-degree count; The node set with zero in-degree count and the activation mark "activated" is determined as the frontier set, and the frontier set and the corresponding adjacency relationship table are saved in ascending order of node number.

[0029] The application realizes explicit modeling of the topological structure and the time sequence logical relationship between tasks by establishing a constraint projection graph according to the time dependence and the causal relationship, and the frontier set determined on this basis enables the system to only process the current effective task node at each stage, thereby significantly reducing the solving scale and realizing accurate control of the causal and priority relationship between tasks, and improving the explainability and topological consistency of the overall scheduling path.

[0030] In the embodiment, the seed selection and the calculation of the composite projection direction specifically include: In the current stage, the seed constraint is selected from the frontier set according to the order rule of priority mark, window index and node number, and the corresponding node number and window index are recorded; According to the directed connection relationship of the constraint projection graph, first-order adjacent constraints are collected, the first-order range includes direct predecessors and direct successors, the adjacent constraint set is formed by truncating according to the preset upper limit, and the node number and the window index of each constraint in the set are recorded. An aggregate weight is calculated for each constraint in the adjacent constraint set, and the aggregate weight is composed of three types of components: a priority component, a stage component, and an adjacent distance component; the three types of components are multiplied to obtain an original weight, and normalization is performed on all original weights to make the weights non-negative and the sum of the weights equal to one; A constraint row vector corresponding to the adjacent constraint is extracted in ascending order of node number, the row vector has the same dimension as the current solution vector, and a mapping table of the row vector sequence and the node number is established; A linear combination of the row vector sequence is performed using the normalized aggregate weight to obtain a composite projection direction, and an index mapping of the composite projection direction and the seed constraint is established; An amplitude clipping is performed to limit the length of the composite projection direction within a predetermined upper limit; A consistency check is performed to compare the composite projection direction with the previous round composite projection direction in terms of directionality, and an angle not greater than a threshold value is considered to pass; if the comparison fails, the aggregate weight is re-normalized and linearly combined again; if it still fails, the seed constraint is replaced and the adjacent collection, weight calculation and linear combination are repeated; After the consistency check passes, the composite projection direction is determined, submitted to the improved Kaczmarz projection update, and the node number, window index, aggregate weight and direction length are recorded for tracing within the stage.

[0031] The present application selects a seed constraint in the front set and aggregates adjacent constraints according to the adjacent relationship to form a composite projection direction, so that the system can adaptively extract global direction information from the local constraint structure, and the normalization of the aggregate weight and the consistency check mechanism can effectively avoid direction shock and convergence deviation in iteration, thereby improving the global stability and direction consistency of the projection update.

[0032] In the embodiment, the execution of the improved Kaczmarz projection update specifically includes: The current solution vector, the composite projection direction, the node number sequence of the adjacent constraint set, the aggregate weight sequence, the constraint row vector sequence and the constant term sequence are read, and the dimension consistency check and the missing value cleaning are completed; The row residual is calculated according to the node number sequence, and the row residual is defined as the inner product of the constant term minus the constraint row vector and the current solution vector, and the row residual sequence is formed and the index correspondence relationship is recorded; The row residual sequence is weighted and summed using the aggregate weight sequence to obtain a weighted residual scalar, and the weighting order is consistent with the node number sequence; The curvature scalar in the composite projection direction is calculated, and the curvature scalar is defined as: the projection length of each constraint row vector in the composite projection direction is squared and weighted and summed according to the aggregate weight, and a positive stable term is added to the summation result; A step length parameter is determined, which is the quotient of the weighted residual scalar and the curvature scalar, and the denominator contains a stabilization term; the step length parameter is set to zero if it is less than zero, and is set to a preset upper limit if it is greater than the preset upper limit; The current solution vector is updated along the composite projection direction according to the step length parameter to generate a candidate solution vector, and the numerical precision and storage order are consistent with the current solution vector; A direction consistency check is performed, which is defined as: the angle between the increment vector obtained by subtracting the current solution vector from the candidate solution vector and the composite projection direction is not greater than a threshold; if not satisfied, the step length parameter is proportionally reduced and updated again until the consistency check is satisfied; if the step length parameter is reduced to the lower limit and still does not satisfy, the update is stopped and the current solution vector is kept unchanged; An amplitude clipping is performed, which is defined as: the Euclidean distance between the candidate solution vector and the current solution vector is not more than a preset radius; if exceeded, the increment vector is proportionally scaled to make the distance equal to the preset radius; The candidate solution vector, the step length parameter, the curvature scalar, the weighted residual scalar, the composite projection direction length, the node number sequence, the aggregation weight sequence, the timestamp and the stage index are recorded as read objects for default integral ledger calculation.

[0033] The present application introduces row residual weighting, curvature scalar calculation and step length self-adaptive adjustment in the improved Kaczmarz projection update process, so that each iteration can complete the solution vector correction within the optimal step length range, realizing the minimization of the constraint error and the high precision control of the update direction, thereby significantly improving the convergence efficiency and scheduling solution accuracy under multiple constraint conditions.

[0034] In the embodiment, the default integral ledger and the budget comparison process specifically includes: A default integral ledger is established for time window measurement, and the ledger fields include a time window index, a budget value, a current integral value, a cumulative integral value, and a timestamp; The candidate solution vector is mapped to the constraint set within the time window, and the constraint row vector and the constant term participating in the calculation are determined according to the node number sequence and the aggregation weight sequence; The row residual is calculated piece by piece, which is defined as the inner product of the constant term minus the constraint row vector and the candidate solution vector, and the row residual sequence is obtained; The positive part of the row residual sequence is taken, which is defined as zero for negative values, and the time window integral increment is obtained by weighted sum according to the aggregation weight sequence in each time window; The current integral value and the cumulative integral value are updated according to the time window, and a snapshot of the node number sequence, the aggregation weight sequence, the stage index and the timestamp is recorded; The current integral value is compared with the budget value piece by piece in each time window, and two types of results, an acceptance mark and a non-acceptance mark, are generated; When there are cases where the label is not accepted, calculate the scaling factor range, select the smallest scaling factor that satisfies all time window budget constraints, and uniformly scale the increment between the candidate solution vector and the current solution vector. After completing the unified scaling, the time window integral increment and comparison results are recalculated. If there are still unacceptable markers, the seed constraint is changed and the composite projection direction is recalculated based on the constraint projection diagram to generate a new candidate solution vector. The mapping, row residual calculation, time window integral increment calculation and comparison are performed again. When all time windows are marked as acceptable, an acceptable update is formed, recording the scaling factor, time window level integral, stage index, timestamp and related index information, and the candidate solution vector is marked as an acceptable update.

[0035] This invention achieves dynamic monitoring and automatic correction of time window constraints and resource budget constraints by establishing a default points ledger measured by time windows and performing budget comparisons. The system can judge the acceptability based on the points results and make scaling adjustments or recalculations, so that the scheduling results maintain a dynamic balance between resource consumption and timing constraints, thereby ensuring the global feasibility and constraint adaptability of the system operation.

[0036] In this embodiment, the acceptance of writes and status updates specifically include: Read the acceptance flag, and simultaneously read the records of candidate solution vector, stage index, timestamp, scaling factor, node number sequence and aggregate weight sequence; When a candidate solution vector is marked as acceptable, it is recorded as the current solution vector, and the stage index and timestamp are registered. In the constraint projection graph and the activation subgraph, mark the seed constraints and the first-order adjacency constraints participating in the composite projection as processed, and remove the processed nodes from the frontier set; Update the in-degree count within the activated subgraph, add nodes with an in-degree count of zero to the frontier set, and maintain the node numbers in ascending order; The stage index is updated based on the homotopy parameter sequence. The stage index is incremented when there are no unprocessed nodes in the active subgraph, and remains unchanged when there are unprocessed nodes in the active subgraph. Write the default points into the default points ledger, which includes time window points, cumulative points, stage index, timestamp, scaling factor, node number sequence and aggregate weight sequence; When the acceptance is marked as non-acceptance, keep the current solution vector unchanged, keep the frontier set unchanged, and record the reason for non-acceptance, the associated time window index and timestamp; Generate a stage state record, which includes the stage index, frontier set, number of active subgraph nodes, acceptance flag, and current solution vector identifier. After recording, proceed to the next iteration.

[0037] The application realizes the orderly control of topology promotion and phase switching in the task scheduling process by synchronously adjusting the frontiers set and the phase index during the receiving write-in and state updating process. The system can maintain the correct evolution of the frontier nodes and the complete record of the constraint ledger after each update, thereby ensuring the logical consistency and global convergence stability of the entire scheduling process.

[0038] Embodiment 1 To verify the feasibility of the application in implementation, the application is applied to the intelligent manufacturing multi-Agent scheduling scene on the cloud-edge-end three-layer cloud native architecture. The infrastructure uses container orchestration and service mesh to provide elasticity and traffic management. The business side exists an order-driven assembly process, including 10 workstations, 35 intelligent Agents, 20 computing nodes and 60 task flows. The historical operation shows that the task arrival presents batch pulse in peak period, the resource occupation is unbalanced, the priority insertion causes queuing fluctuation and energy consumption surge. The traditional scheduling usually relies on fixed priority and periodic recalculation. When facing task dependence and resource causality, it is easy to appear trajectory shock and scheduling rollback. The average resource utilization rate is less than 75%, the energy consumption deviation is higher than 9%, the response time tail is long, and the business side also requires the introduction of skill growth mechanism to make new Agents complete the ability cultivation under the premise of maintaining stable rhythm.

[0039] In the application process, first, the data acquisition and preprocessing module accesses task data, resource data and Agent capability portrait, completes time alignment and standardization, initializes the current solution vector and time sequence constraint set. The time sequence feasible region homotopy module divides the time axis into continuous windows, constructs the homotopy from the initial feasible region to the target feasible region and establishes the phase index. The constraint projection graph module generates a directed graph according to time dependence and causality and determines the frontier set, which is used as the basis for row-level selection. The seed selection and composite projection direction module calculates the composite projection direction in each round of solving. The improved Kaczmarz iteration module performs projection update along the direction to obtain the candidate solution vector. The default integral and budget evaluation module accumulates the integral according to the time window and compares it with the budget. When the mark is acceptable, the receiving write-in and state updating module adopts the candidate solution vector as the current solution vector, and advances the frontier set and the phase index. When it is not acceptable, the recalculation amplitude is scaled or the seed is replaced to recalculate the composite projection direction. The plan generation module outputs task assignment, start and end time, resource allocation and ability cultivation arrangement accordingly, and maintains online iteration when the task flow continues to arrive.

[0040] To reflect the improvement effect, "cloud native priority + rolling recalculation strategy" is selected as the control scheme. The two schemes run for several production cycles under the same task arrival process, the same energy consumption and concurrent limitation, and record the key indicators: average task response time, resource utilization rate, scheduling feasibility, energy consumption deviation rate, rescheduling time consumption and delay rate. The comparison data is shown in Table 1:

[0041] Table 1 Performance comparison table of the system of the application and the traditional cloud native scheduling scheme

[0042] The results show that the application reduces the average task response time by about 35%, the resource utilization rate is increased by about 13 percentage points, the scheduling feasibility is close to the full load interval, the energy deviation rate is reduced to 3.8%, the re-scheduling time is shortened by about 74%, the delay rate is controlled at about 5%, and in combination with the log, it can be seen that the time sequence feasible domain homotopy makes the stage index step more smooth, the constraint projection graph and the frontier set concentrate the row-level selection in the active dependency chain, the improved Kaczmarz converges faster in the composite projection direction, the default integral account book and the budget comparison provide a unified criterion for the adoption of candidate solution vectors, the plan generation continuously outputs executable arrangements in the continuous arrival task stream, and at the same time completes the rhythm allocation of the Agent capability cultivation. The overall results prove that in the cloud-edge-end cloud native background, the application can stably improve the time sequence feasibility, the convergence efficiency and the resource matching accuracy, and significantly reduce the recalculation cost and the energy consumption deviation.

[0043] The above describes only the preferred specific embodiments of the application, but the protection scope of the application is not limited thereto, any person skilled in the art can make equivalent replacements or changes to the technical range disclosed by the application according to the technical scheme and the inventive concept of the application, which should be covered within the protection scope of the application.

Claims

1. An intelligent agent task scheduling and nurturing management system based on time sequence modeling, characterized in that, The method comprises the following steps: a data collection and preprocessing module is used to collect task data, resource data and agent capability profile, perform time alignment and standardization, initialize the current solution vector and the time sequence constraint set; a time sequence feasible region homotopy module is used to generate a windowed time sequence feasible region based on the time sequence constraint set, construct a homotopy from the initial feasible region to the target feasible region, and establish a stage index; a constraint projection graph module is used to construct a constraint projection graph and determine a frontier set according to time dependence and causal relationship; a seed selection and composite projection direction module is used to select a seed constraint from the frontier set in a stage, aggregate adjacent constraints according to adjacency relationship, and calculate a composite projection direction; an improved Kaczmarz iteration module is used to perform improved Kaczmarz projection update along the composite projection direction based on the current solution vector and generate a candidate solution vector; a default integral and budget evaluation module is used to establish a default integral account book measured by time window, calculate the time window integral corresponding to the candidate solution vector and compare it with the budget, generate an acceptance mark, perform update scaling and seed replacement when the acceptance mark is unacceptable, and recalculate the composite projection direction; an acceptance writing and state updating module is used to write the candidate solution vector as the current solution vector when the acceptance mark is acceptable, update the frontier set, the stage index and the default integral account book, and keep the current solution vector unchanged when the acceptance mark is unacceptable; a plan generation module is used to generate task assignment, start and end time, resource allocation and capability development arrangement based on the current solution vector, and output scheduling and development plan.

2. The time-based modeling intelligent Agent task scheduling and nurturing management system according to claim 1, characterized in that, The modules are realized by the following methods: collect task data, resource data and agent capability profile, complete time alignment and standardization, initialize the current solution vector and the time sequence constraint set; generate a windowed time sequence feasible region based on the time sequence constraint set, construct a homotopy from the initial feasible region to the target feasible region, and establish a stage index; construct a constraint projection graph and determine a frontier set according to time dependence and causal relationship; select a seed constraint from the frontier set in the current stage, aggregate adjacent constraints according to adjacency relationship, and calculate a composite projection direction; perform improved Kaczmarz projection update along the composite projection direction based on the current solution vector, and generate a candidate solution vector; establish a default integral account book measured by time window, calculate the time window integral corresponding to the candidate solution vector and compare it with the budget, mark it as acceptable when the time window integral does not exceed the budget, scale the update amplitude when it does not meet the requirement, replace the seed constraint and recalculate the composite projection direction when it still does not meet the requirement; adopt the current solution vector for acceptable update, advance the constraint projection graph frontier and update the stage index and the default integral account book, and keep the current solution vector for unacceptable update and update the default integral account book; generate task assignment, start and end time, resource allocation and capability development arrangement based on the current solution vector, output scheduling and development plan when the termination condition is met, and continue iteration on the updated stage index and frontier set when the termination condition is not met.

3. The time-based modeling intelligent Agent task scheduling and nurturing management system according to claim 2, wherein, The improved Kaczmarz comprises: composite projection direction generation, calculating weights from seed constraints and adjacent constraints based on the constraint projection graph and linearly combining row vectors to form a composite projection direction; Homotopy phase step control, mapping phase index to step size upper limit and direction weight adjustment coefficient and loading to iteration; Half-space projection correction, performing half-space projection on inequality constraints and truncating out-of-bound updates; Budget acceptance criterion, calculating updated time window integral according to default integral account and comparing with budget, scaling update amplitude when exceeding budget, replacing seed constraint and recalculating when still exceeding budget, and accepting update when not exceeding budget; Frontier advancing and phase updating, advancing constraint projection graph frontier and synchronously updating phase index and default integral account when update is accepted; Solution vector writing, recording accepted candidate solution vector as current solution vector and as starting point of next iteration.

4. The time-based modeling intelligent Agent task scheduling and nurturing management system according to claim 2, wherein, The construction of the timing feasible region and homotopy specifically includes: Determining time range and window length, dividing into continuous windows in sequence and numbering, establishing window index set; Assigning timing constraint set to corresponding window according to window index, forming window-level constraint set, and recording time window label, priority label and causality label; Taking window-level constraint set of starting window as basis, constructing initial feasible region, and recording boundary description and variable range; Taking window-level constraint set of target window set as basis, constructing target feasible region, and recording boundary description and variable range; Setting homotopy parameter value sequence, which monotonically increases from zero to one, determining phase number and establishing phase index; In each phase, setting activation state and weight of window-level constraint set according to homotopy parameter, generating corresponding windowed timing feasible region, and synchronously recording phase and window correspondence.

5. The time-based modeling intelligent Agent task scheduling and nurturing management system according to claim 2, wherein, The construction of the constraint projection graph and determination of the frontier set specifically include: According to window index set and timing constraint set, determining time sequence and causality between constraints in sequence, forming relationship determination record; Assigning window index k and row index i to each constraint, setting node label as "k-i", and establishing node list; According to relationship determination record, formulating directed connection rule, limiting that directed connection is established from predecessor to successor when predecessor time is not later than successor and causality trigger exists; Generating directed edge set according to directed connection rule, and together with node list, constituting constraint projection graph; Generating adjacency relation table according to node number, and respectively registering predecessor list and successor list of each node; According to phase index and homotopy parameter, setting activation label of each node in constraint projection graph, with value limited to "activated" and "not activated"; Forming activated subgraph on node with activation label as "activated" and corresponding directed edge, and counting in-degree of each node in activated subgraph as in-degree count; Determining node set with in-degree count as zero and activation label as "activated" as frontier set, and saving frontier set and corresponding adjacency relation table in ascending order according to node number.

6. The time-based modeling intelligent Agent task scheduling and nurturing management system according to claim 2, wherein, The calculation of seed selection and composite projection direction specifically includes: Selecting seed constraint from frontier set according to priority label, window index and node number, and recording node number and window index; According to adjacency relation of constraint projection graph, collecting first-order adjacent constraints having directed connection with seed constraint, setting upper limit of adjacent number and forming adjacent constraint set; An aggregate weight is calculated for each constraint in the set of adjacent constraints, the aggregate weight being generated according to the phase index, the priority mark, and the adjacent distance, and being normalized in the set so that the sum of the weights is one; A constraint row vector corresponding to the adjacent constraint is extracted, the row vector having the same dimension as the current solution vector and being arranged in order according to the node number; The constraint row vector is linearly combined according to the normalized aggregate weight to obtain a composite projection direction, and an index mapping between the seed constraint and the composite projection direction is preserved; The composite projection direction is subjected to amplitude clipping and consistency checking, the amplitude clipping limiting the length of the direction to be less than a preset upper limit, and the consistency checking requiring the direction to be in the same direction as the previous round direction; If the consistency checking fails, the aggregate weight is re-normalized and the composite projection direction is re-calculated, if the consistency checking still fails, the seed constraint is replaced and the adjacent collection, aggregate weight calculation, and linear combination are re-executed, and if the consistency checking passes, the composite projection direction is determined to be used for the improved Kaczmarz projection update.

7. The time-based modeling intelligent Agent task scheduling and nurturing management system according to claim 2, wherein, The execution of the improved Kaczmarz projection update specifically includes: reading the current solution vector, the composite projection direction, the node number sequence of the set of adjacent constraints, the aggregate weight sequence, the constraint row vector, and the constant term; calculating a row residual according to the node number sequence, the row residual being defined as the constant term minus the inner product of the constraint row vector and the current solution vector, and obtaining a row residual sequence; weighting and summing the row residual sequence according to the aggregate weight sequence to obtain a weighted residual scalar; calculating a curvature scalar on the composite projection direction, the curvature scalar being obtained by multiplying the square of the projection length of each constraint row vector on the composite projection direction by the corresponding aggregate weight and then summing the results, and adding a stabilizing term to the sum, the stabilizing term being positive; determining a step parameter, the step parameter being the quotient of the weighted residual scalar and the curvature scalar, the denominator including the stabilizing term; updating the current solution vector along the composite projection direction according to the step parameter to obtain a candidate solution vector; recording the correspondence between the candidate solution vector, the step parameter, the length of the composite projection direction, the node number sequence, and the aggregate weight sequence.

8. The time-based modeling intelligent Agent task scheduling and nurturing management system according to claim 2, wherein, The default integral ledger and budget comparison process specifically includes: establishing a default integral ledger measured by a time window and setting a budget; calculating a time window integral increment according to the candidate solution vector and the constraint set in the time window; updating the current integral with the time window integral increment and comparing it with the budget to generate an acceptance mark and a non-acceptance mark; if there is a non-acceptance mark, uniformly scaling the increments of the candidate solution vector and the current solution vector, recalculating the time window integral increment and the comparison result; if there is still a non-acceptance mark after scaling, replacing the seed constraint and calculating the composite projection direction according to the constraint projection graph, generating a new candidate solution vector and comparing it again; if all time windows are acceptance marks, forming an acceptable update and recording the time window integral and the related index.

9. The time-based modeling intelligent Agent task scheduling and nurturing management system according to claim 2, wherein, The acceptance writing and state updating specifically includes: reading the acceptance mark, the acceptance mark taking the values of acceptable and non-acceptable; if the acceptance mark is acceptable, recording the candidate solution vector as the current solution vector, and recording the phase index and the time stamp; updating the frontier set according to the constraint projection graph and the active subgraph, removing the processed nodes from the frontier set, and adding new nodes with an in-degree of zero to the frontier set; According to the homotopy parameter sequence, the stage index is updated, and the stage index is incremented when there is no unprocessed node in the active subgraph, and the stage index remains unchanged when there is an unprocessed node in the active subgraph; Update the default score ledger, write the time window score, stage index and timestamp; When the acceptance is marked as not accepted, the current solution vector and the frontier set remain unchanged, and the reason for not accepting, the associated time window index and the timestamp are registered.

Images