Conference activity execution task decomposition, arrangement and management method based on AI technology

Abstract

The invention provides a conference activity execution task decomposition, arrangement and management method based on an AI technology, and relates to the technical field of conference management, and the method comprises the steps: obtaining conference demand information, converting the conference demand information into digital representation data, employing a bidirectional recursive decomposition strategy to carry out task decomposition, and generating subtask data; calculating a task similarity matrix and a conflict matrix to construct a task execution directed graph to generate initial arrangement data; collecting an execution state to generate feedback data, calculating a risk assessment score, and adjusting a task execution scheme when the risk assessment score exceeds a threshold value. According to the invention, the execution efficiency of conference activities can be improved, the resource conflict risk is reduced, and dynamic optimization adjustment is realized.

Description

Technical Field

[0001] The present invention relates to conference management technology, and in particular to a method for decomposing, arranging and managing conference activity execution tasks based on AI technology. Background Art

[0002] At present, the organization of conference activities is gradually developing towards digitalization and intelligentization. In particular, it faces increasing complexity in aspects such as multi-task collaboration, resource matching, and process scheduling. The traditional conference activity management method mainly relies on manual experience and static planning, and it is difficult to cope with the dynamic changes of task granularity, the real-time fluctuations of resource status, and the high-frequency interaction relationships between multiple tasks, resulting in low overall arrangement efficiency and unstable execution process.

[0003] With the development of technologies such as artificial intelligence, natural language processing, and graph computing, introducing AI technology into the conference activity management process can realize intelligent assistance throughout the process from conference requirement understanding to task execution. However, existing solutions mostly focus on optimizing a single link, and there is still a lack of a full-process task management mechanism that integrates task modeling, dynamic decomposition, conflict detection, risk feedback, and intelligent adjustment, and cannot effectively meet the real-time scheduling and optimization requirements in complex conference scenarios.

[0004] Especially in the face of large-scale and cross-departmental conference activities, the correlation between tasks is strong and the execution dependencies are complex. Systems driven by static rules are difficult to adapt to the actual situation where tasks change frequently. Therefore, there is an urgent need for a method for decomposing, arranging and managing conference activity execution tasks based on AI technology to achieve accurate understanding of multi-source requirements, intelligent decomposition and conflict optimization of multiple tasks, and improve the intelligent management level and execution efficiency of conference activities. Summary of the Invention

[0005] An embodiment of the present invention provides a method for decomposing, arranging and managing conference activity execution tasks based on AI technology, which can solve the problems in the prior art.

[0006] In the first aspect of the embodiment of the present invention, A method for decomposing, arranging and managing conference activity execution tasks based on AI technology is provided, including: Obtain conference activity requirement information and convert the conference activity requirement information into conference activity digital representation data; Based on the conference activity digital representation data, adopt a two-way recursive decomposition strategy to decompose tasks. The two-way recursive decomposition strategy determines the optimal task decomposition granularity through iterative operations of top-down target decomposition and bottom-up resource recombination, and generates multiple subtask data; Calculate the similarity between multiple subtask data to obtain a task similarity matrix, calculate the correlation entropy between multiple subtask data to obtain a task conflict matrix, combine the task similarity matrix and the task conflict matrix to construct a task execution directed graph, and generate initial task scheduling data based on the task execution directed graph; Collect the execution status of the initial task scheduling data to generate task execution feedback data, calculate the task risk assessment score according to the task execution feedback data. When the task risk assessment score exceeds the preset threshold, generate a task adjustment instruction according to the task status data and resource status data in the task execution feedback data, update the task execution directed graph according to the task adjustment instruction, and generate the final task scheduling data.

[0007] In an optional embodiment, Obtain the meeting activity requirement information, and convert the meeting activity requirement information into meeting activity digital representation data, including: Construct a meeting activity domain knowledge graph, which contains multi-dimensional information nodes and the association relationships between the nodes. Generate a structured requirement collection template based on the meeting activity domain knowledge graph, and obtain the meeting activity requirement information through the structured requirement collection template; Obtain the text content in the meeting activity requirement information, extract the semantic features of the text content, perform sequence annotation on the semantic features, identify the entity and relationship information in the text content, and generate text structured information based on the entity and relationship information; Input the meeting activity requirement information and the text structured information into a multi-head attention network. The multi-head attention network calculates the attention weights based on the information reliability, and performs weighted fusion on the meeting activity requirement information and the text structured information according to the attention weights to obtain the fusion features; Match and map the fusion features with the information nodes in the meeting activity domain knowledge graph to generate the meeting activity digital representation data.

[0008] In an optional embodiment, Based on the meeting activity digital representation data, adopt a bidirectional recursive decomposition strategy to decompose tasks. The bidirectional recursive decomposition strategy determines the optimal task decomposition granularity through iterative operations of top-down goal decomposition and bottom-up resource reorganization, and generates multiple subtask data, including: Construct a task goal tree based on the meeting activity digital representation data. The task goal tree contains multi-level nodes, each node represents a task goal, and there is a hierarchical dependency relationship between the task goals; Based on the hierarchical dependency relationship, adopt the top-down goal decomposition in the bidirectional recursive decomposition strategy to decompose the upper-level task goals into multiple lower-level sub-goals step by step, and calculate the complexity score of each sub-goal to obtain the initial sub-goal set; Calculate the target correlation degree for adjacent sub-goals in the initial sub-goal set, and adopt the bottom-up resource recombination strategy in the bidirectional recursive decomposition strategy to merge sub-goals. Merge adjacent sub-goals with a correlation degree higher than the preset correlation degree threshold into new sub-goals to obtain a recombined sub-goal set; Calculate the resource consumption index for each sub-goal in the recombined sub-goal set, and perform a weighted calculation of the resource consumption index and the complexity score to obtain a task decomposition evaluation index; Perform iterative operations on the initial sub-goal set and the recombined sub-goal set, and optimize based on the task decomposition evaluation index. When the task decomposition evaluation index reaches the optimal value, obtain the optimal task decomposition granularity; Decompose the task objective tree based on the optimal task decomposition granularity, and construct the task objective information and resource requirement information of each decomposed sub-goal node into corresponding sub-task data to generate multiple sub-task data.

[0009] In an alternative embodiment, Calculate the similarity between multiple sub-task data to obtain a task similarity matrix, calculate the correlation entropy between multiple sub-task data to obtain a task conflict matrix, combine the task similarity matrix and the task conflict matrix to construct a task execution directed graph, and generate initial task scheduling data based on the task execution directed graph, including: Extract the task type, objective, and resource requirement information from multiple sub-task data, generate an attribute vector for each sub-task, calculate the cosine similarity of the attribute vectors of any two sub-tasks, and construct a task similarity matrix; Calculate the time overlap degree by comparing the execution time intervals of any two sub-tasks, calculate the resource competition degree through resource allocation conflict detection, calculate the task dependence strength through task precedence relationship analysis, normalize the time overlap degree, resource competition degree, and task dependence strength, and calculate the task correlation entropy to construct a task conflict matrix; Perform an adaptive weighted combination of the task similarity matrix and the task conflict matrix, dynamically adjust the similarity weight and conflict weight according to the task urgency, generate a combined weight matrix, construct a task execution directed graph based on the combined weight matrix, calculate the sum of the out-degree weights of each vertex in the task execution directed graph, and calculate the transfer influence of the vertex according to the combined weight with adjacent vertices. Combine the sum of the out-degree weights and the transfer influence to obtain the dynamic priority score of the vertex; Select the vertex with the highest dynamic priority score as the starting node, traverse the task execution directed graph, recalculate the dynamic priority scores of the remaining vertices after traversing each vertex, and determine the next traversed vertex according to the updated priority scores to generate initial task scheduling data.

[0010] In an alternative embodiment, Calculate the time overlap degree by comparing the execution time intervals of any two subtasks, calculate the resource competition degree through resource allocation conflict detection, and calculate the task dependence strength through task precedence relationship analysis, including: Obtain the execution time intervals of multiple subtasks, calculate the time fluctuation range of each subtask according to historical execution data, determine the time elasticity coefficient based on the time fluctuation range, and combine the time elasticity coefficient with the execution time interval to construct a fuzzy time window; Perform probability density modeling on the fuzzy time windows of any two subtasks, generate a time point sequence through the Monte Carlo sampling method, calculate the task overlap probability of each time point, and multiply and accumulate the task overlap probability by the time interval of the sampling time point to obtain the time overlap degree; Scan and count the resource application lists of each subtask to obtain resource application frequency data, calculate the usage frequency of resources according to the resource application frequency data; calculate the resource scarcity degree based on the ratio of the total resource amount to the total application amount; multiply the usage frequency by the resource scarcity degree to obtain the occupancy intensity weight of the resource type; Based on the occupancy intensity weight, calculate the product of the demand quantities of any two subtasks on the same resource type, multiply the product of the demand quantities by the occupancy intensity weight to obtain the weighted resource demand quantity; divide the weighted resource demand quantity by the total resource demand quantities of the two subtasks to obtain the resource competition degree; Construct a task dependence directed graph, determine the direct dependence relationship between task nodes through data flow analysis and set it as the edge weight, calculate the influence propagation path in a recursive manner based on the edge weight, and combine the edge weight of the direct dependence relationship with the attenuation weight on the influence propagation path to obtain the task dependence strength.

[0011] In an alternative embodiment, Collect the execution status of the initial task scheduling data to generate task execution feedback data. Calculate the task risk assessment score according to the task execution feedback data, including: Collect the execution status of the initial task scheduling data, dynamically adjust the sampling time interval based on the state change frequency of the task execution node, and group the collected execution status according to an adaptive time window; calculate the fluctuation variance of the execution status for each time window, automatically adjust the time attenuation weight according to the fluctuation variance, and smooth the execution status based on the time attenuation weight to generate task execution feedback data; Construct a multi-layer risk state propagation network, where the bottom-layer nodes of the network correspond to risk factors, the middle-layer nodes represent risk clustering categories, and the top-layer nodes represent the comprehensive risk level; calculate the risk propagation coefficient between nodes based on the task execution feedback data, and update the states of each layer of nodes through bottom-up risk accumulation based on the risk propagation coefficient to obtain the initial risk state, and then correct the states of each layer of nodes through top-down constraint propagation to generate a risk state vector including task progress risk, resource utilization risk, and quality compliance risk; Combine the risk state vector with the task execution feedback data for feature combination, calculate the correlation coefficient matrix between features, assign combined weights to each feature based on the correlation coefficient matrix, and weighted fuse the risk state vector and the task execution feedback data according to the combined weights to generate a task risk assessment score.

[0012] In an alternative embodiment, Generate a task adjustment instruction according to the task state data and resource status data in the task execution feedback data, and update the task execution directed graph according to the task adjustment instruction to generate the final task scheduling data, including: Combine the task state data and resource status data in the task execution feedback data with the task risk assessment score to form a state input, and construct a task priority adjustment policy network and a resource allocation policy network based on the state input; the task priority adjustment policy network generates a task priority adjustment coefficient according to the task state data, and the resource allocation policy network generates a resource allocation ratio according to the resource status data; Store the state input, task priority adjustment coefficient, resource allocation ratio, and their corresponding task risk assessment scores in the experience replay pool; train the policy network based on the experience replay pool, optimize the network parameters according to the change of the task risk assessment score, and form a task adjustment instruction with the optimized task priority adjustment coefficient and resource allocation ratio; In the task execution directed graph, identify the task nodes that need to be adjusted based on the task priority adjustment coefficient in the task adjustment instruction; calculate the influence degree between nodes according to the task state data and resource status data; construct a subgraph to be adjusted with the node set whose influence degree exceeds the influence degree threshold; Establish a constraint network within the subgraph to be adjusted, update the dependency relationship between nodes based on the task state data and resource status data, adjust the task priority and resource allocation according to the task adjustment instruction, and update the task execution directed graph; Store the task adjustment instruction and the task execution feedback data as a training sample; retrieve similar samples based on the current task state data and resource status data, continuously optimize the policy network, and generate the final task scheduling data.

[0013] In the second aspect of the embodiments of the present invention, Provide an electronic device, including: A processor; A memory for storing instructions executable by the processor; Wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0014] In the third aspect of the embodiments of the present invention, A computer-readable storage medium is provided, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0015] In this embodiment, a bidirectional recursive decomposition strategy is adopted for task decomposition, realizing an organic combination of top-down goal decomposition and bottom-up resource recombination, ensuring the scientificity and rationality of task decomposition, and avoiding problems of resource mismatch and low execution efficiency caused by traditional unidirectional task decomposition. By constructing a task similarity matrix and a task conflict matrix to generate a task execution directed graph, the dependency relationships and execution orders between tasks are systematically presented, providing a clear path guidance for the execution of conference activities, and effectively improving the organization efficiency and quality of conference activities. Based on the task execution feedback data, real-time assessment and dynamic adjustment of task risks are realized. When the risk assessment score exceeds the preset threshold, the task execution plan can be updated in a timely manner, enhancing the risk response ability and resource allocation flexibility during the execution of conference activities, and ensuring the smooth progress of conference activities. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is a schematic flowchart of a method for task decomposition, scheduling and management of conference activity execution based on AI technology in the embodiments of the present invention; Figure 2 It is a relationship diagram between the task combination weight matrix and the dynamic priority score in the embodiments of the present invention; Figure 3 It is a probability density distribution diagram of time overlap in the embodiments of the present invention; Figure 4 It is a schematic structural diagram of the task execution directed graph in the embodiments of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0017] To make the objectives, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0018] The technical solution of the present invention will be described in detail below with specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments.

[0019] Figure 1 As shown in the flowchart of the method for task decomposition, orchestration and management of conference activity execution tasks based on AI technology in the embodiments of the present invention, Figure 1 as shown, the method includes: Obtain the conference activity requirement information, and convert the conference activity requirement information into digital representation data of the conference activity; Based on the digital representation data of the conference activity, use a bidirectional recursive decomposition strategy to decompose tasks. The bidirectional recursive decomposition strategy determines the optimal task decomposition granularity through top-down goal decomposition and bottom-up resource reorganization iterative operations, and generates multiple subtask data; Calculate the similarity between multiple subtask data to obtain a task similarity matrix, calculate the correlation entropy between multiple subtask data to obtain a task conflict matrix, combine the task similarity matrix and the task conflict matrix to construct a task execution directed graph, and generate initial task orchestration data based on the task execution directed graph; Collect the execution status of the initial task orchestration data to generate task execution feedback data, calculate the task risk assessment score according to the task execution feedback data. When the task risk assessment score exceeds the preset threshold, generate a task adjustment instruction according to the task status data and resource status data in the task execution feedback data, update the task execution directed graph according to the task adjustment instruction, and generate the final task orchestration data.

[0020] In an alternative embodiment, Obtaining the conference activity requirement information and converting the conference activity requirement information into digital representation data of the conference activity includes: Construct a knowledge graph for the conference activity field. The knowledge graph for the conference activity field contains multi-dimensional information nodes and the association relationships between the nodes. Generate a structured requirement collection template based on the knowledge graph for the conference activity field, and obtain the conference activity requirement information through the structured requirement collection template; Obtain the text content in the conference activity requirement information, extract the semantic features of the text content, perform sequence annotation on the semantic features, identify the entity and relationship information in the text content, and generate text structured information based on the entity and relationship information; Input the conference activity requirement information and the text structured information into a multi-head attention network. The multi-head attention network calculates the attention weights based on the information reliability, and performs weighted fusion on the conference activity requirement information and the text structured information according to the attention weights to obtain a fusion feature; Match and map the fusion features with the information nodes in the knowledge graph of the conference and event domain to generate digital representation data of the conference and event.

[0021] Exemplarily, first construct a knowledge graph of the conference and event domain. This knowledge graph contains multi-dimensional information nodes related to conferences and events and the association relationships between the nodes. The multi-dimensional information nodes include basic information nodes such as conference type nodes, participant nodes, conference venue nodes, conference equipment nodes, conference time nodes, etc., as well as derivative information nodes such as conference process nodes, conference agenda nodes, conference material nodes, etc. The association relationships between the nodes include inheritance relationships, composition relationships, dependency relationships, etc. For example, there is an inheritance relationship between the international academic conference node and the conference type node, and a dependency relationship between the conference agenda node and the conference time node. The various dimensional features of the conference and event and their interactions can be described through the nodes and relationships in the knowledge graph.

[0022] Generate a structured requirement collection template based on the constructed knowledge graph of the conference and event domain. This template sets collection items according to the node hierarchy of the knowledge graph, and the collection items are organized through the association relationships in the knowledge graph. Each collection item in the collection template includes attribute information such as collection item name, data type, value range, and whether it is required. For example, for the conference type collection item, its data type is an enumeration type, and the value range includes preset conference type options such as academic conferences, business conferences, training conferences, etc. Through this template, the requirement information of the conference and event can be obtained in a standardized manner. Specifically, the collection template includes multiple collection areas such as the basic information collection area, venue requirement collection area, personnel requirement collection area, equipment requirement collection area, process requirement collection area, etc. Among them, the basic information collection area is used to collect basic information such as conference name, conference time, and conference location; the venue requirement collection area is used to collect venue requirement information such as venue capacity and venue layout; the personnel requirement collection area is used to collect personnel requirement information such as the number of participants and special guests; the equipment requirement collection area is used to collect equipment requirement information such as projection equipment and audio equipment; the process requirement collection area is used to collect process requirement information such as conference agenda and conference materials.

[0023] After obtaining the text content in the meeting activity requirement information, it is necessary to extract semantic features and perform sequence labeling on the text content. First, use the character vector model to convert the text content into a word vector sequence, and then extract the context semantic features of the text through a bidirectional long short-term memory network. Specifically, when implementing, the character vector model uses a pre-trained Chinese word vector model, which is trained through a large amount of Chinese corpora and can map Chinese characters to a low-dimensional dense vector space. The bidirectional long short-term memory network contains recurrent neural network layers in both the forward and backward directions, and can capture the forward and backward context information of the text at the same time. For example, for the text content "A meeting room that can accommodate 200 people", first convert it into a word vector sequence, and then extract the semantic features including the venue capacity requirement of this text through the bidirectional long short-term memory network.

[0024] When performing sequence labeling on the extracted semantic features, a conditional random field model is used to identify entity and relationship information in the text. Specifically, input the semantic features into the conditional random field model, which calculates the labeling probability based on the feature template and outputs the optimal labeling sequence. The labeling categories include entity labels such as meeting type labels, venue labels, equipment labels, personnel labels, etc., and relationship labels such as venue capacity relationship, equipment quantity relationship, etc. For example, for the above text, the conditional random field model labels "meeting room" as a venue entity and labels the relationship between "200 people" and "meeting room" as a venue capacity relationship. Based on the labeling results, the unstructured text content can be converted into structured information containing entities and relationships.

[0025] Input the meeting activity requirement information and the text structured information into a multi-head attention network for information fusion. The multi-head attention network contains multiple attention sub-networks, and each attention sub-network independently calculates the attention scores between the query vector, key vector, and value vector. The query vector is generated from the meeting activity requirement information, and the key vector and value vector are generated from the text structured information. The attention score represents the degree of association between different information. Specifically, when implementing, first calculate the information reliability. The information with higher reliability obtains a higher attention weight during fusion. The information reliability is comprehensively evaluated based on multiple dimensions such as information source, information integrity, and information consistency. For example, the information reliability collected through a structured template is higher than the information extracted from unstructured text. Then each attention sub-network calculates the attention weight based on the information reliability and performs weighted summation on the meeting activity requirement information and the text structured information to obtain the fusion feature.

[0026] Finally, the fused features are matched and mapped with the information nodes in the knowledge graph of the conference activity field. Specifically, the similarity between the fused features and each information node in the knowledge graph is calculated, and the information node with the highest similarity is selected as the mapping target. The similarity calculation is based on the attribute features and relationship features of the nodes. The attribute features include the attribute values, attribute types, etc. of the nodes, and the relationship features include the adjacent nodes, relationship types, etc. of the nodes. For example, if the fused features contain the attribute value of "academic conference", then it has a high similarity with the academic conference node in the knowledge graph. Through feature mapping, the fused features can be converted into a normalized representation based on the knowledge graph, generating digital representation data of the conference activity. This digital representation data contains various dimensional features of the conference activity and realizes the associated expression of features through the organizational structure of the knowledge graph.

[0027] In this embodiment, a structured requirement collection template is constructed through the knowledge graph of the conference activity field to obtain conference activity requirement information in a standardized manner, improving the integrity and accuracy of the requirement information; the bidirectional long short-term memory network and the conditional random field model are used to extract semantic features and perform sequence annotation on the text content to accurately identify entity and relationship information in the text; the multi-head attention network is used to perform feature fusion based on information reliability to effectively integrate conference activity information from different sources and improve the accuracy of information fusion; the normalized representation of conference activity information is achieved through node matching and mapping of the knowledge graph, making the digital representation data have good structure and interpretability. This method can convert unstructured conference activity requirement information into normalized digital representation data, providing a reliable data basis for subsequent intelligent analysis and management of conference activities.

[0028] In an alternative embodiment, Based on the digital representation data of the conference activity, a bidirectional recursive decomposition strategy is adopted for task decomposition. The bidirectional recursive decomposition strategy determines the optimal task decomposition granularity through iterative operations of top-down goal decomposition and bottom-up resource reorganization, generating multiple subtask data including: Construct a task goal tree based on the digital representation data of the conference activity. The task goal tree contains multi-level nodes, and each node represents a task goal. There is a hierarchical dependency relationship between task goals; Based on the hierarchical dependency relationship, the top-down goal decomposition in the bidirectional recursive decomposition strategy is used to gradually decompose the upper-level task goals into multiple lower-level sub-goals, and the complexity score of each sub-goal is calculated to obtain an initial set of sub-goals; Calculate the goal correlation degree between adjacent sub-goals in the initial set of sub-goals, and use the bottom-up resource reorganization strategy in the bidirectional recursive decomposition strategy to merge sub-goals. The adjacent sub-goals with a correlation degree higher than the preset correlation degree threshold are merged into new sub-goals to obtain a set of reorganized sub-goals; Calculate the resource consumption index for each sub-goal in the set of recombinant sub-goals, and perform a weighted calculation of the resource consumption index and the complexity score to obtain a task decomposition evaluation index; Perform iterative operations on the initial sub-goal set and the recombinant sub-goal set, optimize based on the task decomposition evaluation index, and when the task decomposition evaluation index reaches the optimal value, obtain the optimal task decomposition granularity; Decompose the task goal tree based on the optimal task decomposition granularity, and construct the task goal information and resource requirement information of each sub-goal node obtained by the decomposition into corresponding sub-task data, generating multiple sub-task data.

[0029] Exemplarily, construct a task goal tree based on the digital representation data of the conference activity. This task goal tree adopts a tree-like hierarchical structure and contains multiple levels of task goal nodes. The task goal nodes form a hierarchical dependency relationship through the parent-child relationship. The upper-level nodes represent relatively macroscopic task goals, and the lower-level nodes represent specific execution goals. Each task goal node contains task goal information and resource requirement information. The task goal information describes the specific goal content that this node needs to complete, and the resource requirement information describes the resource requirements such as personnel, equipment, and venue required to complete this goal. For example, for an academic conference, the root node can be "hold an academic seminar", and its lower-level nodes include task goal nodes such as "conference registration", "conference agenda arrangement", and "on-site service". These nodes can be further decomposed into more specific sub-task goal nodes.

[0030] Adopt the top-down goal decomposition method in the bidirectional recursive decomposition strategy, and based on the hierarchical dependency relationship in the task goal tree, gradually decompose the upper-level task goal nodes into multiple lower-level sub-goal nodes. The decomposition process first analyzes the constituent elements of the upper-level task goal, identifies the key sub-goals required to complete this goal, and then maps these sub-goals to the lower-level task goal nodes. Calculate the complexity score for each sub-goal node. The complexity score is comprehensively evaluated based on multiple dimensions such as the time cost, labor cost, and technical difficulty required to complete the goal. Specifically in implementation, set weight coefficients for each dimension, and sum the weighted scores of each dimension to obtain the final complexity score. For example, "conference agenda arrangement" can be decomposed into sub-goals such as "formulate the conference schedule", "arrange the speech order", and "coordinate the venue time". Among them, "coordinate the venue time" has a relatively high complexity score due to involving multi-party communication and coordination. Obtain the initial sub-goal set through recursive decomposition.

[0031] Calculate the target correlation degree for adjacent sub-goals in the initial sub-goal set. The target correlation degree represents the degree of association between two sub-goals. The higher the correlation degree, the more suitable the two sub-goals are for combined processing. The target correlation degree is calculated based on the resource similarity, execution dependency, and timing overlap degree of the sub-goals. Resource similarity represents the coincidence degree of the resources required by two sub-goals. Execution dependency represents the strength of the execution dependency relationship between two sub-goals. Timing overlap degree represents the degree of overlap of the execution times of two sub-goals. For example, the two sub-goals of "arranging the speech order" and "coordinating the venue time" have a high resource similarity and execution dependency, and their target correlation degree is high.

[0032] Adopt the bottom-up resource recombination strategy in the bidirectional recursive decomposition strategy to merge adjacent sub-goals with a target correlation degree higher than the preset correlation degree threshold. In the merging process, first select a pair of adjacent sub-goals with the highest target correlation degree, merge them into a new sub-goal, and then recalculate the correlation degree between the new sub-goal and its adjacent sub-goals. Repeat the merging operation until there are no sub-goal pairs that meet the merging conditions. The new sub-goal after merging inherits the resource requirement information of the original sub-goal and integrates the overlapping resources. For example, merge "arranging the speech order" and "coordinating the venue time" into the new sub-goal of "speech time arrangement and coordination". Obtain the recombined sub-goal set through resource recombination.

[0033] Calculate the resource consumption index for each sub-goal in the recombined sub-goal set. The resource consumption index reflects the total amount of resources required to complete the sub-goal, including multiple dimensions such as human resource consumption, equipment resource consumption, and venue resource consumption. The resource consumption in each dimension is quantitatively calculated based on the resource requirement information of the sub-goal. Perform a weighted calculation on the resource consumption index and the previously calculated complexity score to obtain the task decomposition evaluation index. This evaluation index comprehensively reflects the rationality of the task decomposition scheme. The smaller the evaluation index, the better the decomposition scheme.

[0034] Iteratively optimize the initial sub-goal set and the recombined sub-goal set. In each round of iteration, alternately perform top-down goal decomposition and bottom-up resource recombination, and calculate the corresponding task decomposition evaluation index. When the relative change rate of the task decomposition evaluation index is less than the preset threshold, it is considered to reach the optimal state, and determine the sub-goal division scheme of the current iteration round as the optimal task decomposition granularity. Decompose the task goal tree based on the optimal task decomposition granularity, and extract the task goal information and resource requirement information in the decomposed sub-goal nodes. Organize the extracted information according to the predefined data structure to construct sub-task data including target description, resource configuration, etc. For example, for the sub-goal of "speech time arrangement and coordination", the generated sub-task data includes specific time arrangement requirements, required coordinating personnel, and other information. Finally, output multiple structured sub-task data for subsequent task assignment and execution management.

[0035] In this embodiment, through the hierarchical representation of the task objective tree, the hierarchical dependency relationship between task objectives is clearly described; a two-way recursive decomposition strategy combining top-down objective decomposition and bottom-up resource recombination is adopted, which not only ensures the integrity of task decomposition but also realizes the optimization of resource utilization; the task decomposition evaluation index based on the objective complexity score and resource consumption index can effectively evaluate the rationality of the task decomposition scheme; the optimal task decomposition granularity is determined through iterative optimization, achieving a balance between the degree of task objective refinement and resource allocation efficiency; the finally generated subtask data contains both clear objective descriptions and detailed resource configuration information, providing a reliable basis for subsequent task execution. This method can adaptively disassemble complex conference activity tasks into a reasonable set of subtasks, improving task execution efficiency and resource utilization rate.

[0036] In an alternative embodiment, Calculate the similarity between multiple subtask data to obtain a task similarity matrix, calculate the correlation entropy between multiple subtask data to obtain a task conflict matrix, combine the task similarity matrix and the task conflict matrix to construct a task execution directed graph, and generate initial task scheduling data based on the task execution directed graph, including: Extract the task type, objective, and resource requirement information from multiple subtask data, generate an attribute vector for each subtask, calculate the cosine similarity between the attribute vectors of any two subtasks, and construct a task similarity matrix; Calculate the time overlap degree by comparing the execution time intervals of any two subtasks, calculate the resource competition degree through resource allocation conflict detection, calculate the task dependence intensity through task precedence relationship analysis, normalize the time overlap degree, resource competition degree, and task dependence intensity, and then calculate the task correlation entropy to construct a task conflict matrix; Perform an adaptive weighted combination of the task similarity matrix and the task conflict matrix, dynamically adjust the similarity weight and conflict weight according to the task urgency, generate a combined weight matrix, construct a task execution directed graph based on the combined weight matrix, calculate the sum of the out-degree weights of each vertex in the task execution directed graph, and calculate the transfer influence of the vertex according to the combined weight with adjacent vertices. Combine the sum of the out-degree weights and the transfer influence to obtain the dynamic priority score of the vertex; Select the vertex with the highest dynamic priority score as the starting node, traverse the task execution directed graph, recalculate the dynamic priority scores of the remaining vertices after traversing each vertex, and determine the next traversed vertex according to the updated priority scores to generate initial task scheduling data.

[0037] Exemplarily, task type, task objective, and resource requirement information are extracted from multiple subtask data to generate an attribute vector for each subtask. The task type information includes the nature category of the task, such as meeting preparation, meeting site, meeting service, etc.; the task objective information includes the specific objective description and completion criteria of the task; the resource requirement information includes the required personnel type, equipment type, venue type, etc. These information are converted into an attribute vector in a unified format, and each dimension in the attribute vector corresponds to a task feature. For example, for the "conference registration service" task, its attribute vector contains information such as conference service type, registration personnel requirement, registration equipment requirement, etc.

[0038] The cosine similarity is calculated for the attribute vectors of any two subtasks to obtain the similarity degree between these two tasks. The cosine similarity measures the similarity of task features by calculating the cosine value of the angle between two attribute vectors. The larger the similarity value, the more similar the two tasks are. The similarity values between all task pairs are organized into a task similarity matrix, which reflects the similarity relationship between each task in the task set. For example, the two tasks of "conference registration service" and "participant information collection" have a relatively high similarity, while the similarity with the "venue arrangement" task is relatively low.

[0039] The time overlap degree is calculated by comparing the execution time intervals of any two subtasks. The time overlap degree represents the degree of overlap in time between two tasks and is measured by calculating the proportion of the intersection of the execution time intervals of the two tasks to the union. The larger the time overlap degree, the more likely the two tasks are to have a time conflict during execution. For example, there is partial overlap in the execution times of the two tasks of "conference registration service" and "venue arrangement", and the execution order needs to be reasonably arranged.

[0040] The resource competition degree is calculated through resource allocation conflict detection. The resource competition degree represents the degree of conflict in resource usage between two tasks and is calculated by detecting whether there is an overlap in the resource requirements of the two tasks. For each resource type, if both tasks need to use this type of resource and the total amount of resources is limited, there is a resource competition. The larger the resource competition degree, the more likely the two tasks are to have a resource conflict during execution. For example, there is a relatively high resource competition degree between two tasks that need to use the same meeting room.

[0041] The task dependency strength is calculated through task precedence relationship analysis. The task dependency strength represents the execution dependency relationship between two tasks and is calculated by analyzing the input-output relationship of the tasks. If the output of one task is a necessary input for another task, there is a dependency relationship between these two tasks. The larger the dependency strength, the less the execution order of the two tasks can be adjusted. For example, "participant information collection" must be completed before "conference material preparation", and there is a strong dependency relationship between these two tasks.

[0042] Normalize the time overlap degree, resource competition degree, and task dependence strength so that their value ranges are unified between zero and one, and then calculate the task correlation entropy. The task correlation entropy comprehensively reflects the conflict degree between two tasks during execution. The larger the correlation entropy, the more serious the conflict between tasks. Organize the correlation entropy values between all task pairs into a task conflict matrix.

[0043] Perform an adaptive weighted combination of the task similarity matrix and the task conflict matrix. Dynamically adjust the similarity weight and conflict weight according to the urgency of the task. For urgent tasks, give priority to considering conflict factors, while for non-urgent tasks, consider similarity factors more. Obtain a combined weight matrix through weight combination, and construct a task execution directed graph based on this matrix. Each vertex in the directed graph represents a task, and the directed edge between vertices represents the execution order between tasks. The weight of the edge is determined by the combined weight.

[0044] Calculate the sum of the out-degree weights of each vertex in the task execution directed graph. The sum of the out-degree weights represents the direct association strength between this task and other tasks. At the same time, calculate the transfer influence of the vertex. The transfer influence evaluates the indirect influence range of the task by considering the combined weights with adjacent vertices. Combine the sum of the out-degree weights and the transfer influence to obtain the dynamic priority score of the vertex. This score comprehensively reflects the importance of the task in the entire task network.

[0045] Select the vertex with the highest dynamic priority score as the starting node for traversal. This is usually the most influential or urgent task. During the traversal process, after visiting each vertex, recalculate the dynamic priority scores of the remaining vertices. The update of the priority score takes into account the influence of the visited vertices on the remaining tasks to ensure the dynamic adaptability of task scheduling. Determine the next traversal vertex according to the updated priority score, and repeat this process until all vertices are traversed, finally generating the initial task scheduling data. This scheduling data reflects the execution order and priority arrangement of tasks.

[0046] Figure 2 This is the relationship diagram between the task combined weight matrix and the dynamic priority score in the embodiment of the present invention, as Figure 2As shown, the figure shows a directed graph of task execution containing 8 task nodes and 12 connection edges, as well as a corresponding comparison table of task priority scores. Each node in the figure represents a subtask, and the thickness of the connection lines between nodes intuitively represents the magnitude of the combined weight. Among them, the combined weight between Task 1 and Task 2 is the highest, reaching 0.82, while the combined weight between Task 1 and Task 5 is the lowest, only 0.23. These combined weights are obtained through the adaptive weighted combination of the task similarity matrix and the task conflict matrix, fully considering the similarity and potential conflicts between tasks. In terms of dynamic priority scores, Task 2 has the highest score of 0.95 because its out-degree weight sum (the sum of the combined weights with Task 3 and Task 6) is large and its transmission influence is strong; while Task 8 has the lowest score of 0.52, mainly because it is at the end of the task chain and its transmission influence is limited. By comparing the dynamic priority of this technical solution with traditional static priority sorting methods (such as the EDD algorithm based on the deadline, the full name is Earliest Due Date), it can be found that the priority of this technical solution is generally higher than that of the static method, and it can better capture the impact of the dependency relationship between tasks on the overall execution efficiency. For example, for the key node Task 2, the priority given by this solution is 0.95, significantly higher than about 0.80 of the static method, which reflects that this solution can identify and strengthen the importance of key tasks. It can also be seen from the figure that the directed graph of task execution presents an obvious network structure, reflecting the complex interdependent relationship between tasks in the actual working environment. Compared with the traditional linear or tree-like task structure, the directed graph constructed by this technical solution can express the multi-dimensional association between tasks more comprehensively.

[0047] By constructing a task similarity matrix and a task conflict matrix and performing an adaptive weighted combination, this application establishes a directed graph of task execution and generates initial task scheduling data based on a dynamic priority mechanism, enabling accurate modeling and dynamic sorting of the collaborative relationship and conflict relationship between multiple subtasks. Existing technologies mostly adopt static priorities or fixed rules for allocation in task scheduling, making it difficult to take into account both the relevance and conflict of tasks at the same time, and easily leading to resource conflicts, execution blockages or low scheduling efficiency. The starting point for the improvement of this application is to consider the dynamic balance between the collaborative feasibility reflected by task similarity and the execution risk represented by task conflict. An attribute vector is constructed by integrating task type, target and resource requirement information to quantify the similarity between tasks. At the same time, time overlap, resource competition and pre-order relationship are introduced to evaluate the degree of task conflict. Different factors are fused through an adaptive weighting method, and the combined weight is dynamically adjusted in combination with the urgency of the task, making the task scheduling more flexible and real-time. Compared with traditional methods, this solution can not only effectively avoid the allocation of conflicting resources, but also improve the overall task execution efficiency and rationality, and has higher scheduling flexibility and execution stability.

[0048] In an alternative embodiment, the time overlap degree is calculated by comparing the execution time intervals of any two subtasks, the resource competition degree is calculated by resource allocation conflict detection, and the task dependency strength is calculated by task precedence relationship analysis, including: Obtain the execution time intervals of multiple subtasks, calculate the time fluctuation range of each subtask according to historical execution data, determine the time elasticity coefficient based on the time fluctuation range, and combine the time elasticity coefficient with the execution time interval to construct a fuzzy time window; Perform probability density modeling on the fuzzy time windows of any two subtasks, generate a time point sequence through the Monte Carlo sampling method, calculate the task overlap probability of each time point, and multiply and accumulate the task overlap probability by the time interval of the sampling time point to obtain the time overlap degree; Scan and count the resource application lists of each subtask to obtain resource application frequency data, calculate the usage frequency of resources according to the resource application frequency data; calculate the resource scarcity degree based on the ratio of the total resource amount to the total application amount; multiply the usage frequency by the resource scarcity degree to obtain the occupancy intensity weight of the resource type; Based on the occupancy intensity weight, calculate the product of the demand quantities of any two subtasks on the same resource type, multiply the product of the demand quantities by the occupancy intensity weight to obtain the weighted resource demand quantity; divide the weighted resource demand quantity by the total resource demand quantities of the two subtasks to obtain the resource competition degree; Construct a task dependency directed graph, determine the direct dependency relationship between task nodes through data flow analysis and set it as the edge weight, calculate the influence propagation path in a recursive manner based on the edge weight, and combine the edge weight of the direct dependency relationship with the attenuation weight on the influence propagation path to obtain the task dependency strength.

[0049] This embodiment provides a task coordination analysis method based on subtask time intervals, resource occupancy, and dependency relationships. By calculating the time overlap degree, resource competition degree, and task dependency strength between tasks, efficient scheduling and optimized execution of tasks are achieved.

[0050] First, obtain the execution time interval data of multiple subtasks. Taking four subtasks A, B, C, and D as an example, extract the execution time interval data from the historical execution records. The execution time interval of subtask A is [10, 20], indicating that the start time is at the 10th moment and the end time is at the 20th moment; the execution time interval of subtask B is [15, 25]; the execution time interval of subtask C is [5, 15]; the execution time interval of subtask D is [22, 30].

[0051] Calculate the time fluctuation range of each subtask based on historical execution data. For subtask A, by collecting its historical execution records, it is found that its actual execution time fluctuates within the range of ±10% of the nominal value. Therefore, the time fluctuation range is ±2; the time fluctuation range of subtask B is ±3; the time fluctuation range of subtask C is ±1.5; the time fluctuation range of subtask D is ±2.4.

[0052] Determine the time elasticity coefficient according to the time fluctuation range. The calculation method of the time elasticity coefficient is the time fluctuation range divided by the nominal execution time length. For subtask A, the time elasticity coefficient is 2 / 10 = 0.2; for subtask B it is 0.3; for subtask C it is 0.15; for subtask D it is 0.3.

[0053] Combine the time elasticity coefficient with the execution time interval to construct a fuzzy time window. The fuzzy time window of subtask A is [8, 22], which is obtained by shifting the original start time forward by 2 units and the end time backward by 2 units; the fuzzy time window of subtask B is [12, 28]; the fuzzy time window of subtask C is [3.5, 16.5]; the fuzzy time window of subtask D is [19.6, 32.4].

[0054] Perform probability density modeling on the fuzzy time windows of any two subtasks. The fuzzy time window indicates that the actual execution possibility of the task is distributed within a certain interval, and the boundaries are not completely determined but have a certain degree of ambiguity. Probability density modeling is used for the modeling of the fuzzy time window to quantify the possibility distribution of task execution at different time points. Taking subtasks A and B as an example, assign a probability density value to each time point within the fuzzy time window. The central time point has the highest probability, which decreases towards both ends. Using the triangular probability density function, the probability density of subtask A at time point 10 is 0.1, at time point 15 is 0.3, and at time point 20 is 0.1; the probability density of subtask B at time point 15 is 0.08, at time point 20 is 0.25, and at time point 25 is 0.08.

[0055] Generate a time point sequence through the Monte Carlo sampling method. Based on the above probability density function, randomly generate 1000 sampling points within the fuzzy time window range, including the start time and end time points. For subtask A, some of the generated time sampling points are [8.2, 9.5, 10.8, 12.4, 14.7, 15.9, 17.3, 18.6, 20.2, 21.8]; for subtask B, some of the generated time sampling points are [12.3, 13.8, 15.2, 16.7, 18.4, 20.1, 21.9, 23.5, 25.3, 27.8].

[0056] Calculate the task overlap probability for each time point. For any time point t, if this time point falls within the fuzzy time windows of two subtasks simultaneously, then calculate the probability that the two subtasks are executed simultaneously at this time point, that is, the product of the probability density values of the two tasks at this time point. At time point 16, the probability density of subtask A is 0.25, and the probability density of subtask B is 0.15, then the overlap probability is 0.0375.

[0057] Multiply the task overlap probability by the time interval of the sampling time points and accumulate to obtain the time overlap degree. The time overlap degree is used to measure the conflict risk degree of two tasks in the execution time. The larger the value, the higher the possibility that the two tasks conflict in time. The time interval of the sampling time points is set to 0.1, then the time overlap degree of subtasks A and B is 0.0375×0.1 + 0.0412×0.1 +... + 0.0298×0.1 = 0.421, indicating that these two subtasks overlap approximately 42.1% in time.

[0058] Scan and count the resource application lists of each subtask to obtain the resource application frequency data. Assume that the system includes three types of resources: CPU, memory, and disk. Subtask A applies for 5 units of CPU resources, 10 units of memory resources, and 20 units of disk resources; Subtask B applies for 8 units of CPU resources, 5 units of memory resources, and 15 units of disk resources. Count the resource application lists of all subtasks to obtain that the total number of times the CPU resources are applied is 25 times, the total number of times the memory resources are applied is 35 times, and the total number of times the disk resources are applied is 80 times.

[0059] Calculate the usage frequency of resources according to the resource application frequency data. The usage frequency of CPU resources is 25 / 140 = 0.179, the usage frequency of memory resources is 35 / 140 = 0.25, and the usage frequency of disk resources is 80 / 140 = 0.571.

[0060] Calculate the resource scarcity degree based on the ratio of the total resource amount to the total application amount. The total amount of CPU resources is 20 units, the total application amount is 25 units, and the resource scarcity degree is 25 / 20 = 1.25; the total amount of memory resources is 40 units, the total application amount is 35 units, and the resource scarcity degree is 35 / 40 = 0.875; the total amount of disk resources is 100 units, the total application amount is 80 units, and the resource scarcity degree is 80 / 100 = 0.8.

[0061] Multiply the usage frequency by the resource scarcity degree to obtain the occupancy intensity weight of the resource type. The occupancy intensity weight of CPU resources is 0.179×1.25 = 0.224; the occupancy intensity weight of memory resources is 0.25×0.875 = 0.219; the occupancy intensity weight of disk resources is 0.571×0.8 = 0.457.

[0062] Based on the occupancy intensity weights, calculate the product of the requirements of any two subtasks on the same resource type, and multiply the product of the requirements by the occupancy intensity weights to obtain the weighted resource requirements. The product of the requirements of subtasks A and B on the CPU is 5 × 8 = 40, and the weighted resource requirement is 40 × 0.224 = 8.96; the product of the requirements on the memory is 10 × 5 = 50, and the weighted resource requirement is 50 × 0.219 = 10.95; the product of the requirements on the disk is 20 × 15 = 300, and the weighted resource requirement is 300 × 0.457 = 137.1.

[0063] Divide the weighted resource requirements by the total resource requirements of the two subtasks to obtain the resource competition degree. The total resource requirements of subtasks A and B are (5 + 10 + 20) + (8 + 5 + 15) = 63, and the resource competition degree between the two is (8.96 + 10.95 + 137.1) / 63 = 2.49.

[0064] Construct a task dependency directed graph, and determine the direct dependency relationship between task nodes through data flow analysis and set it as the edge weight. Assume that 60% of the output data of subtask A is used by subtask B, then set the edge weight from A to B to 0.6; 30% of the output data of subtask B is used by subtask D, then set the edge weight from B to D to 0.3; 40% of the output data of subtask C is used by subtask A, then set the edge weight from C to A to 0.4; 25% of the output data of subtask C is used by subtask B, then set the edge weight from C to B to 0.25.

[0065] Based on the edge weights, calculate the impact propagation path in a recursive manner. For the impact propagation path from C to D, there are two paths: C → A → B → D and C → B → D. The weight of the C → A → B → D path is 0.4 × 0.6 × 0.3 = 0.072; the weight of the C → B → D path is 0.25 × 0.3 = 0.075.

[0066] Combine the edge weights of the direct dependency relationship with the attenuation weights on the impact propagation path to obtain the task dependency intensity. Use an attenuation factor of 0.8, that is, the impact of the indirect dependency relationship decays as the path length increases. The task dependency intensity from subtask C to D is 0.072 × 0.8² + 0.075 × 0.8 = 0.106, indicating that subtask C has a 10.6% dependency impact on subtask D.

[0067] Figure 3 This is the time overlap probability density distribution diagram of the embodiment of the present invention, as Figure 3As shown, this figure shows the probability distribution of the time overlap between tasks T2 and T4. Through the probability density modeling of the fuzzy time window and the Monte Carlo sampling method, this technical solution obtains a more realistic time overlap probability distribution curve. It can be clearly seen from the figure that the probability density distribution calculated by this technical solution shows an asymmetric bimodal structure. The probability density at 10 hours is 0.224, and it reaches the highest value of 0.241 at 15 hours, which reflects the complex time relationship between the two tasks during execution. In contrast, the linear probability distribution method (such as the linear attenuation model) shows a relatively gentle unimodal distribution, with the highest point only being 0.195 at 16 hours; while the uniform probability distribution method (such as the standard PERT method) simply assumes that the probability density is constant at 0.143 within the 0-20 hour interval and cannot capture the dynamic characteristics of time overlap. At the 20-hour point, the probability density given by this technical solution is 0.150, which is higher than that of the uniform distribution but lower than that of the linear distribution, indicating that this solution can more accurately identify the overlap characteristics at the time window boundary. The advantage of this technical solution is that it considers the combined effects of the time fluctuation range and the elasticity coefficient, and the probability density generated by a large number of sampling points is closer to the statistical law of actual task execution. This accurate probability density modeling is crucial for accurately calculating the time overlap degree, enabling the system to better predict and handle time conflicts during task execution, thereby improving the overall scheduling efficiency.

[0068] In this embodiment, by introducing the fuzzy time window and the probability modeling method, the time overlap risk between tasks is accurately quantified, effectively improving the accuracy of time conflict recognition; by integrating the resource usage frequency and scarcity to evaluate the resource occupancy intensity, a fine-grained modeling of the resource competition relationship is achieved; by combining the data flow dependency analysis and the attenuation mechanism of the recursive propagation path, the direct and indirect dependency relationships between tasks are accurately characterized. Compared with the prior art that often uses static configuration or rough rules to handle task conflict problems and cannot balance the temporal flexibility and resource coordination between tasks, this solution takes the dynamic feature modeling as the core, takes into account the execution uncertainty and system resource constraints, constructs a multi-dimensional conflict evaluation mechanism, realizes the comprehensive quantitative analysis of the scheduling risk, and effectively improves the accuracy, robustness of task scheduling and the overall operation efficiency of the system.

[0069] In an alternative implementation, collecting the execution status of the initial task scheduling data to generate task execution feedback data, and calculating the task risk assessment score according to the task execution feedback data includes: Collect the execution status of the initial task scheduling data, dynamically adjust the sampling time interval based on the status change frequency of the task execution nodes, and group the collected execution status according to an adaptive time window; calculate the fluctuation variance of the execution status for each time window, automatically adjust the time decay weight according to the fluctuation variance, and smooth the execution status based on the time decay weight to generate task execution feedback data; Construct a multi-layer risk status propagation network, where the bottom-layer nodes of the network correspond to risk factors, the middle-layer nodes represent risk clustering categories, and the top-layer nodes represent the comprehensive risk level; calculate the risk propagation coefficient between nodes based on the task execution feedback data, and update the status of each layer of nodes through bottom-up risk accumulation based on the risk propagation coefficient to obtain the initial risk status, and then correct the status of each layer of nodes through top-down constraint propagation to generate a risk status vector including task progress risk, resource utilization risk, and quality compliance risk; Combine the risk status vector with the task execution feedback data for feature combination, calculate the correlation coefficient matrix between features, assign a combined weight to each feature based on the correlation coefficient matrix, and weighted fuse the risk status vector and the task execution feedback data according to the combined weight to generate a task risk assessment score.

[0070] This embodiment provides a task risk assessment method, including collecting the execution status of the initial task scheduling data to generate task execution feedback data, and calculating the task risk assessment score according to the task execution feedback data.

[0071] Exemplarily, the system collects the execution status of the initial task scheduling data. This process dynamically adjusts the sampling time interval based on the status change frequency of the task execution nodes to achieve a balance between sampling efficiency and data accuracy. Specifically, the system monitors the frequency value of the status change of the execution node. When it detects that the status change frequency exceeds a preset threshold (for example, the number of status changes per minute is greater than 5 times), the system automatically shortens the sampling interval (such as shortening from the original 60 seconds to 30 seconds); conversely, when the status change frequency is lower than another preset threshold (for example, the number of status changes per minute is less than 1 time), the system extends the sampling interval (such as extending from the original 30 seconds to 60 seconds). This adaptive sampling strategy can effectively reduce data redundancy while ensuring the capture of key status change points.

[0072] Group the collected execution status according to an adaptive time window. The size of the time window is dynamically adjusted according to the task complexity and execution stage. For example, for development tasks, a smaller window (e.g., 2 hours) can be set for the coding stage, and a larger window (e.g., 4 hours) for the testing stage. For each time window, the system calculates the variance of the execution status fluctuations, which represents the stability of the execution status within that time window. In practical applications, the variance of fluctuations can be obtained by calculating the statistical dispersion of the status indicators (such as CPU utilization, memory occupancy, task completion percentage, etc.) within the window.

[0073] Based on the calculated variance of fluctuations, the system automatically adjusts the time decay weights. The larger the variance of fluctuations, the more unstable the status, and higher weights are assigned to recent data; the smaller the variance of fluctuations, the relatively more stable the status, and a more balanced weight distribution is adopted. For example, when the variance of fluctuations exceeds 0.3, the weights of the data in the most recent 30% time period can be set to 0.6, and the weights of the data in the remaining 70% time period can be set to 0.4; when the variance of fluctuations is below 0.1, the weights can be set to 0.5 and 0.5 respectively to achieve balanced weighting. Based on these time decay weights, the system smooths the execution status to generate task execution feedback data. The smoothing process can effectively filter out the interference of short-term fluctuations and retain the main trend of status changes.

[0074] After generating the task execution feedback data, the system constructs a multi-layer risk status propagation network for task risk assessment. This network adopts a hierarchical structure, consisting of bottom-layer, middle-layer, and top-layer nodes. The bottom-layer nodes correspond to specific risk factors, such as insufficient resources, schedule delay, code defects, etc.; the middle-layer nodes represent risk clustering categories, such as resource risk, schedule risk, quality risk; the top-layer node represents the comprehensive risk level, reflecting the overall risk status of the task.

[0075] Based on the task execution feedback data, the system calculates the risk propagation coefficients between the nodes in the network. The risk propagation coefficient reflects the strength of the influence relationship between different risk factors, with a value range from 0 to 1, and the larger the value, the stronger the influence. For example, in a software development task, the propagation coefficient of "insufficient testing resources" on "decline in code quality" may be 0.7, indicating that insufficient testing resources have a strong impact on code quality; while the propagation coefficient of "unstable development environment" on "schedule delay" may be 0.3, indicating a relatively weak influence.

[0076] After the risk propagation coefficient is determined, the system updates the status of each layer of nodes through a bottom-up risk accumulation process to obtain the initial risk status. Specifically, starting from the bottom-layer risk factor nodes, the risk status value is multiplied by the corresponding propagation coefficient and accumulated and transmitted to the upper-layer nodes, and the calculation is performed layer by layer until the top-layer nodes. Subsequently, the system corrects the status of each layer of nodes through a top-down constraint propagation process, that is, the top-layer comprehensive risk information is reversely transmitted to the lower-layer nodes to correct the initial risk assessment, so that the risk assessment result is more comprehensive and accurate.

[0077] Through this two-way propagation process, the system generates a risk status vector including task progress risk, resource utilization risk, and quality compliance risk. For example, the risk status vector of a software project may be expressed as [0.6, 0.3, 0.7], corresponding to the degrees of progress risk, resource risk, and quality risk respectively. The larger the value, the higher the risk.

[0078] Finally, the system combines the risk status vector with the task execution feedback data and calculates the correlation coefficient matrix between the features. The correlation coefficient matrix reflects the association strength between various risk factors and execution indicators, which helps to identify key risk indicators. Based on the correlation coefficient matrix, the system assigns combined weights to each feature. The stronger the correlation between features, the higher the assigned weight. For example, for the product R & D task of an enterprise, if the correlation coefficient between "code submission frequency" and "quality compliance risk" is 0.8, then in the quality risk assessment, this feature may obtain a weight of 0.3; while the correlation coefficient between "team communication frequency" and "progress risk" is 0.5, then it may obtain a weight of 0.2 in the progress risk assessment.

[0079] According to the determined combined weights, the system weights and fuses the risk status vector with the task execution feedback data to generate the final task risk assessment score. This risk assessment score is usually represented by a value from 0 to 100. The higher the score, the greater the risk. For example, the risk assessment score of a R & D project is 75, indicating that the project has a high risk and requires key monitoring and intervention; the score of another project is 25, indicating a low risk and the existing management strategy can be maintained.

[0080] In this embodiment, through the dynamic sampling and time window grouping method, it is possible to accurately adapt to the change frequency of the task state, avoid the problems of information lag or data redundancy caused by a fixed sampling interval, and achieve efficient monitoring of the task execution state. At the same time, through the time decay weight mechanism driven by the fluctuation variance, the adaptive smoothing of the state data is realized, and the interference of accidental fluctuations on risk judgment is effectively suppressed. By using a multi-layer risk state propagation network to structurally model the task risk, it is possible to start from multiple risk sources, comprehensively analyze the potential risks in dimensions such as task progress, resource utilization, and quality control, and achieve hierarchical expression and dynamic correction of risk characteristics. Through the feature correlation analysis and weight fusion mechanism, the comprehensiveness and accuracy of task risk assessment are effectively improved. Compared with the existing technology that only evaluates task risks based on a single indicator or a static model, this solution can more accurately identify potential abnormal trends in task execution, thus significantly improving the risk perception ability and response ability of the scheduling system.

[0081] In an alternative embodiment, a task adjustment instruction is generated according to the task state data and resource status data in the task execution feedback data, and the task execution directed graph is updated according to the task adjustment instruction. The generation of the final task scheduling data includes: The task state data and resource status data in the task execution feedback data are combined with the task risk assessment score to form a state input. A task priority adjustment policy network and a resource allocation policy network are constructed based on the state input; the task priority adjustment policy network generates a task priority adjustment coefficient according to the task state data, and the resource allocation policy network generates a resource allocation ratio according to the resource status data; The state input, task priority adjustment coefficient, resource allocation ratio, and their corresponding task risk assessment scores are stored in the experience replay pool; the policy network is trained based on the experience replay pool, the network parameters are optimized according to the change of the task risk assessment score, and the optimized task priority adjustment coefficient and resource allocation ratio are combined to form a task adjustment instruction; In the task execution directed graph, the task nodes that need to be adjusted are identified based on the task priority adjustment coefficient in the task adjustment instruction; the influence degree between nodes is calculated according to the task state data and resource status data; a sub-graph to be adjusted is constructed for the node set whose influence degree exceeds the influence degree threshold; A constraint network is established within the sub-graph to be adjusted, the dependency relationship between nodes is updated based on the task state data and resource status data, the task priority and resource allocation are adjusted according to the task adjustment instruction, and the task execution directed graph is updated; The task adjustment instruction and the task execution feedback data are combined to form a training sample for storage; similar samples are retrieved based on the current task state data and resource status data, and the policy network is continuously optimized to generate the final task scheduling data.

[0082] This embodiment provides an adaptive task scheduling method, which dynamically adjusts task priorities and resource allocations based on task execution feedback data to improve the overall execution efficiency of the system.

[0083] Exemplarily, task status data, resource status data, and task risk assessment scores are extracted from the task execution feedback data to construct a status input. The task status data includes the completion degree, execution duration, and dependency completion status of each task; the resource status data includes the computing resource utilization rate, network bandwidth occupancy, and storage space consumption; the task risk assessment score is calculated based on historical task execution data and the current system load.

[0084] For example, for a distributed computing task, the system may collect the following status data: the completion degree of task A is 70%, and the execution duration has exceeded the expected by 30%; the completion degree of task B is 35%, and the execution duration meets the expectation; the CPU utilization rate is 89%, the memory occupancy is 65%, and the network bandwidth occupancy is 42%; the risk assessment score of task A is 0.78, and the risk assessment score of task B is 0.45.

[0085] Based on the status input, the system constructs a task priority adjustment policy network and a resource allocation policy network. The task priority adjustment policy network consists of an input layer, two hidden layers, and an output layer. The input layer receives the task status data, and the output layer generates the priority adjustment coefficients for each task. The hidden layer processes the features using the ReLU activation function and focuses on key task status indicators through the attention mechanism. In the above example, this network may output a priority adjustment coefficient of +0.25 (increase the priority) for task A and a coefficient of -0.15 (decrease the priority) for task B.

[0086] The resource allocation policy network has a similar structure but focuses on processing the resource status data and outputs the allocation ratios for each resource type. This network uses the Softmax function to ensure that the sum of all resource allocation ratios is 1. Continuing with the above example, this network may allocate 70% of the CPU resources and 60% of the memory resources to task A, and 30% of the CPU resources and 40% of the memory resources to task B.

[0087] The system stores the status input, task priority adjustment coefficients, resource allocation ratios, and the corresponding task risk assessment scores in an experience replay pool. The experience replay pool uses a key-value storage structure and uses a hash index to accelerate the retrieval of similar states. Each stored record contains the state vector, the adjustment policy taken, and the change in the risk score after adjustment.

[0088] For the data in the experience replay pool, the system uses the temporal difference learning algorithm to train the policy network. Specifically, the system randomly extracts a batch of experience samples, calculates the difference between the expected risk score under the current policy and the change in the actual risk score, and updates the network parameters through the backpropagation algorithm. The optimization goal is to minimize the task risk assessment score, that is, to improve the task execution success rate and reduce resource waste. After multiple rounds of training, the network parameters gradually converge, generating a better task priority adjustment coefficient and resource allocation ratio, which form a task adjustment instruction.

[0089] After the task adjustment instruction is generated, the system identifies the task nodes that need to be adjusted in the task execution directed graph. Specifically, the system traverses the task execution directed graph and selects the nodes whose absolute value of the priority adjustment coefficient exceeds the adjustment threshold (e.g., 0.2) as the adjustment targets. For the above example, the adjustment coefficient of task A is +0.25, which exceeds the threshold, so it is marked as needing adjustment. Next, the system calculates the influence degree between nodes. The influence degree is calculated by considering the direct dependency relationship, the degree of shared resources, and the historical execution correlation between nodes. The system uses the sliding window algorithm to analyze the recent task execution data and identifies strongly correlated task pairs. For example, if the execution status of task C is highly correlated with task A (correlation coefficient 0.85, exceeding the influence degree threshold 0.7), then task C is also included in the range to be adjusted. The system constructs these nodes into a subgraph to be adjusted.

[0090] Within the subgraph to be adjusted, the system builds a constraint network and updates the dependency relationship between nodes. The constraint network is represented using a graph data structure, where nodes represent tasks and edges represent dependency relationships or resource constraints. The system updates the weight of the edges according to the latest task status data to reflect the current dependency strength. For example, if it is found that a new data dependency has been added between task A and task D, the system will add an edge from task D to task A in the constraint network, with the weight set to 0.9, indicating a strong dependency relationship.

[0091] The system adjusts the priorities and resource allocations of the tasks in the subgraph to be adjusted according to the task adjustment instruction. For tasks with increased priorities (such as task A), the system advances them to the earliest execution position allowed by the dependencies through the topological sorting algorithm; for resource allocation adjustments, the system modifies the allocation parameters of the resource scheduler to ensure that high-priority tasks obtain sufficient resources. After the update is completed, the system synchronizes the changes to the task execution directed graph.

[0092] During the implementation process, the system stores the task adjustment instructions and task execution feedback data for each time as training samples. These samples include the state before adjustment, the adjustment actions taken, and the system performance after adjustment. The system uses the K-nearest neighbor algorithm to retrieve similar historical samples based on the current task status data and resource status data, analyzes the historical adjustment effects, and further optimizes the current policy network parameters.

[0093] The system generates task scheduling data based on the updated task execution directed graph, including the execution order, priority, resource allocation scheme, and estimated completion time of each task. The task scheduling data is output in JSON format and distributed to the task scheduler for execution. For example, the final scheduling data may show that the priority of task A is increased from 3 to 4, the CPU resource allocation is increased from 50% to 70%, and the estimated completion time is shortened by 25%.

[0094] Figure 4 FIG. 1 is a schematic diagram of a directed graph structure of task execution according to an embodiment of the present invention. Figure 4 As shown in the figure, the complete task dependency structure and critical path are shown. The figure contains 16 task nodes, distributed on 6 levels, and the dependencies between tasks are represented by directed edges. Among them, Task 3, Task 7, Task 11, Task 13, Task 14 and Task 16 (circular nodes) constitute the critical path, with a total execution time of 87.5 minutes and a total impact of 4.61. Each task node is marked with its priority value, ranging from 0.4 (Task 10) to 0.95 (Task 16). Nodes on the critical path generally have a higher priority. The impact between tasks is clearly marked next to the connecting line, among which Task 1→Task 3 has the highest impact of 0.95, Task 3→Task 7 is 0.92, Task 11→Task 14 is 0.91, and Task 14→Task 16 is 0.94. These high-impact connections generally appear on the critical path. The subgraph to be adjusted is marked with a dotted box in the figure, which contains a set of nodes whose impact exceeds the threshold of 0.85. These nodes will be included in the priority adjustment range. Through this graphical representation, the system can intuitively identify the bottleneck location and optimization space in the task execution process, so as to apply the task priority adjustment coefficient and resource allocation ratio in a targeted manner to improve the overall execution efficiency. Compared with the traditional static task scheduling, this dynamic adjustment method can more accurately respond to state changes and resource fluctuations during the execution process.

[0095] In this embodiment, by constructing a task priority adjustment policy network and a resource allocation policy network, it is possible to dynamically generate task adjustment instructions according to the real-time task status and resource conditions, realizing intelligent adjustment and optimization in the task scheduling process. Compared with the existing methods that generally use rule-driven or static models for task priority sorting and resource allocation, which cannot effectively cope with the risk fluctuations and resource bottlenecks that occur during task execution, this application adopts a data-driven policy network mechanism, continuously optimizing the network parameters through an experience replay pool to improve the adaptability and robustness of the adjustment instructions. Further, by constructing a subgraph to be adjusted and introducing a method for analyzing the degree of influence and updating the constraint network, it is possible to accurately identify the key adjustment areas and effectively control the adjustment range, avoiding the chain effects caused by large changes in the global task structure. Finally, continuous training and updating of the policy network are achieved in the feedback loop, significantly enhancing the system's response ability and self-evolution ability in a dynamic environment, and improving the intelligent level of task orchestration and the full-process scheduling efficiency.

[0096] In the second aspect of the embodiments of the present invention, a kind of electronic device is provided, including: a processor; a memory for storing instructions executable by the processor; wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0097] In the third aspect of the embodiments of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0098] The present invention can be a method, device, system and / or computer program product. The computer program product can include a computer-readable storage medium, on which computer-readable program instructions for executing various aspects of the present invention are uploaded.

[0099] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for decomposing, scheduling, and managing meeting activity execution tasks based on AI technology, characterized in that, Including: Obtain the meeting activity requirement information and convert the meeting activity requirement information into digital representation data of the meeting activity; Based on the digital representation data of the meeting activity, adopt a bidirectional recursive decomposition strategy to decompose tasks. The bidirectional recursive decomposition strategy determines the optimal task decomposition granularity through iterative operations of top-down goal decomposition and bottom-up resource recombination, and generates multiple subtask data; Calculate the similarity between multiple subtask data to obtain a task similarity matrix, calculate the correlation entropy between multiple subtask data to obtain a task conflict matrix, combine the task similarity matrix and the task conflict matrix to construct a task execution directed graph, and generate initial task scheduling data based on the task execution directed graph; Collect the execution status of the initial task scheduling data to generate task execution feedback data, calculate the task risk assessment score according to the task execution feedback data. When the task risk assessment score exceeds the preset threshold, generate a task adjustment instruction according to the task status data and resource status data in the task execution feedback data, update the task execution directed graph according to the task adjustment instruction, and generate the final task scheduling data.

2. The method according to claim 1, wherein Obtain the meeting activity requirement information and convert the meeting activity requirement information into digital representation data of the meeting activity, including: Construct a knowledge graph for the meeting activity field. The knowledge graph for the meeting activity field contains multi-dimensional information nodes and the association relationships between the nodes. Generate a structured requirement collection template based on the knowledge graph for the meeting activity field, and obtain the meeting activity requirement information through the structured requirement collection template; Obtain the text content in the meeting activity requirement information, extract the semantic features of the text content, perform sequence annotation on the semantic features, identify the entity and relationship information in the text content, and generate text structured information based on the entity and relationship information; Input the meeting activity requirement information and the text structured information into a multi-head attention network. The multi-head attention network calculates the attention weights based on the information reliability, and performs weighted fusion on the meeting activity requirement information and the text structured information according to the attention weights to obtain a fused feature; Match and map the fused feature with the information nodes in the knowledge graph for the meeting activity field to generate digital representation data of the meeting activity.

3. The method according to claim 1, wherein Based on the digital representation data of the meeting activity, adopt a bidirectional recursive decomposition strategy to decompose tasks. The bidirectional recursive decomposition strategy determines the optimal task decomposition granularity through iterative operations of top-down goal decomposition and bottom-up resource recombination, and generates multiple subtask data, including: Construct a task goal tree based on the digital representation data of the meeting activity. The task goal tree contains multi-level nodes, each node represents a task goal, and there is a hierarchical dependency relationship between the task goals; Based on the hierarchical dependency relationship, adopt the top-down goal decomposition in the bidirectional recursive decomposition strategy to gradually decompose the upper-level task goals into multiple lower-level sub-goals, and calculate the complexity score of each sub-goal to obtain an initial sub-goal set; Calculate the goal correlation degree between adjacent sub-goals in the initial sub-goal set, adopt the bottom-up resource recombination strategy in the bidirectional recursive decomposition strategy to merge sub-goals, and merge adjacent sub-goals with a correlation degree higher than the preset correlation degree threshold into new sub-goals to obtain a recombined sub-goal set; Calculate the resource consumption index for each sub-goal in the recombinant sub-goal set, and perform weighted calculation on the resource consumption index and the complexity score to obtain the task decomposition evaluation index; Perform iterative operations on the initial sub-goal set and the recombinant sub-goal set, optimize based on the task decomposition evaluation index, and when the task decomposition evaluation index reaches the optimal value, obtain the optimal task decomposition granularity; Decompose the task goal tree based on the optimal task decomposition granularity, and construct the corresponding sub-task data from the task goal information and resource requirement information of each decomposed sub-goal node to generate multiple sub-task data.

4. The method according to claim 1, wherein Calculate the similarity between multiple sub-task data to obtain the task similarity matrix, calculate the correlation entropy between multiple sub-task data to obtain the task conflict matrix, combine the task similarity matrix and the task conflict matrix to construct a task execution directed graph, and generate initial task scheduling data based on the task execution directed graph, including: Extract the task type, goal, and resource requirement information from multiple sub-task data, generate the attribute vector of each sub-task, calculate the cosine similarity of the attribute vectors of any two sub-tasks, and construct the task similarity matrix; Calculate the time overlap degree by comparing the execution time intervals of any two sub-tasks, calculate the resource competition degree through resource allocation conflict detection, calculate the task dependence strength through task precedence relationship analysis, normalize the time overlap degree, resource competition degree, and task dependence strength, and calculate the task correlation entropy to construct the task conflict matrix; Perform adaptive weighted combination on the task similarity matrix and the task conflict matrix, dynamically adjust the similarity weight and conflict weight according to the task urgency, generate a combined weight matrix, construct a task execution directed graph based on the combined weight matrix, calculate the sum of the out-degree weights of each vertex in the task execution directed graph, and calculate the transfer influence of the vertex according to the combined weight with the adjacent vertex, and combine the sum of the out-degree weights and the transfer influence to obtain the dynamic priority score of the vertex; Select the vertex with the highest dynamic priority score as the starting node, traverse the task execution directed graph, recalculate the dynamic priority scores of the remaining vertices after traversing each vertex, determine the next traversed vertex according to the updated priority scores, and generate the initial task scheduling data.

5. The method according to claim 4, wherein Calculate the time overlap degree by comparing the execution time intervals of any two sub-tasks, calculate the resource competition degree through resource allocation conflict detection, calculate the task dependence strength through task precedence relationship analysis, including: Obtain the execution time intervals of multiple sub-tasks, calculate the time fluctuation range of each sub-task according to historical execution data, determine the time elasticity coefficient based on the time fluctuation range, and combine the time elasticity coefficient and the execution time interval to construct a fuzzy time window; Perform probability density modeling on the fuzzy time windows of any two sub-tasks, generate a time point sequence through the Monte Carlo sampling method, calculate the task overlap probability of each time point, and multiply and accumulate the task overlap probability by the time interval of the sampling time point to obtain the time overlap degree; Scan and count the resource application lists for each subtask to obtain resource application frequency data, calculate the usage frequency of resources according to the resource application frequency data; calculate the resource scarcity degree based on the ratio of the total resource quantity to the total application quantity; multiply the usage frequency by the resource scarcity degree to obtain the occupancy intensity weight of the resource type. Based on the occupancy intensity weight, calculate the product of the demand quantities of any two subtasks for the same resource type, multiply the product of the demand quantities by the occupancy intensity weight to obtain the weighted resource demand quantity; divide the weighted resource demand quantity by the total resource demand quantities of the two subtasks to obtain the resource competition degree. Construct a task dependency directed graph, determine the direct dependency relationships between task nodes through data flow analysis and set them as edge weights, calculate the impact propagation paths in a recursive manner based on the edge weights, and combine the edge weights of the direct dependency relationships with the attenuation weights on the impact propagation paths to obtain the task dependency intensity.

6. The method according to claim 1, wherein Collect the execution status of the initial task scheduling data to generate task execution feedback data, calculate the task risk assessment score according to the task execution feedback data, including: Collect the execution status of the initial task scheduling data, dynamically adjust the sampling time interval based on the state change frequency of the task execution nodes, group the collected execution status according to an adaptive time window; calculate the fluctuation variance of the execution status for each time window, automatically adjust the time decay weight according to the fluctuation variance, and smooth the execution status based on the time decay weight to generate task execution feedback data. Construct a multi-layer risk state propagation network, where the bottom layer nodes correspond to risk factors, the middle layer nodes represent risk clustering categories, and the top layer nodes represent the comprehensive risk degree; calculate the risk propagation coefficients between nodes based on the task execution feedback data, update the states of each layer of nodes through bottom-up risk accumulation based on the risk propagation coefficients to obtain the initial risk state, and then correct the states of each layer of nodes through top-down constraint propagation to generate a risk state vector including task progress risk, resource utilization risk, and quality compliance risk. Combine the risk state vector with the task execution feedback data, calculate the correlation coefficient matrix between features, assign combination weights to each feature based on the correlation coefficient matrix, and weighted fuse the risk state vector and the task execution feedback data according to the combination weights to generate the task risk assessment score.

7. The method according to claim 1, wherein Generate task adjustment instructions according to the task status data and resource status data in the task execution feedback data, update the task execution directed graph according to the task adjustment instructions, and generate the final task scheduling data, including: Form a state input by combining the task status data and resource status data in the task execution feedback data with the task risk assessment score, and construct a task priority adjustment strategy network and a resource allocation strategy network based on the state input; the task priority adjustment strategy network generates a task priority adjustment coefficient according to the task status data, and the resource allocation strategy network generates a resource allocation ratio according to the resource status data. Store the status input, task priority adjustment coefficient, resource allocation ratio, and their corresponding task risk assessment scores in the experience replay pool; train the policy network based on the experience replay pool, optimize the network parameters according to the change of the task risk assessment score, and form a task adjustment instruction with the optimized task priority adjustment coefficient and resource allocation ratio; In the task execution directed graph, identify the task nodes to be adjusted based on the task priority adjustment coefficient in the task adjustment instruction; calculate the influence degree between nodes according to the task status data and resource status data; construct a subgraph to be adjusted with the set of nodes whose influence degree exceeds the influence degree threshold; Establish a constraint network within the subgraph to be adjusted, update the dependency relationship between nodes based on the task status data and resource status data, adjust the task priority and resource allocation according to the task adjustment instruction, and update the task execution directed graph; Store the task adjustment instruction and task execution feedback data as training samples; retrieve similar samples based on the current task status data and resource status data, continuously optimize the policy network, and generate the final task scheduling data.

8. An electronic device, characterized in that, Comprising: A processor; A memory for storing instructions executable by the processor; Wherein, the processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 7.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, the method according to any one of claims 1 to 7 is implemented.

Images