A multi-skill agent scheduling method based on a skill capability relationship graph

Abstract

The application discloses a multi-skill agent scheduling method based on a skill capability relation graph, first, a skill capability relation graph containing skill nodes and dependency, mutual exclusion and hierarchical correlation is constructed; then, a user task is parsed to extract a capability label, skill semantic correlation and comprehensive score are calculated, and candidate skills are screened; based on the graph constraint, a path search is completed to generate an optimal skill execution path, a gradual skill disclosure mechanism is adopted to guide a language model to execute a task; finally, skill execution data is recorded and a historical success rate is dynamically updated, and a scheduling strategy is iteratively optimized. The application realizes structured modeling of skill scheduling, effectively reduces a decision space, improves skill selection accuracy and system stability, has a continuous self-optimization capability, and is suitable for complex multi-skill collaborative task scenes.

Description

Technical Field

[0001] This invention relates to the technical field of multi-skill intelligent agents driven by large language models, and in particular to a multi-skill intelligent agent scheduling method based on skill ability relationship graphs. Background Technology

[0002] With the development of large language models, language model-driven multi-skill intelligent agent systems have gradually become an important technical means to automate complex tasks. In such systems, different domain capabilities are usually encapsulated in the form of skills. Each skill contains prompts, knowledge descriptions, and tool interfaces to guide the language model in completing a specific task, thereby achieving modularity and scalability of system capabilities.

[0003] In existing technologies, agents typically select appropriate skills to perform tasks by semantically matching skill descriptions using language models based on user input. However, as the number of skills in a system continues to increase and task complexity continues to rise, existing technologies are gradually revealing the following shortcomings:

[0004] First, an increase in the number of skills leads to a sharp expansion of the model's decision space. When there are many skills, the language model needs to match and judge among a large number of candidate skills, which can easily lead to problems such as semantic overlap and functional intersection, resulting in a decrease in skill selection accuracy, instability of inference paths, and an increase in inference costs.

[0005] Secondly, the existing skill selection process mainly relies on implicit reasoning of language models and lacks an explicit structured decision-making mechanism, which makes the skill selection process difficult to quantify and interpret, and unable to effectively constrain the skill matching results, thereby reducing the stability and controllability of the system.

[0006] Furthermore, in existing technologies, skills typically exist as independent capability modules, lacking a modeling mechanism for the dependencies, hierarchical relationships, and mutual exclusions between skills, making it difficult for the system to perform reasonable skill scheduling and priority ranking.

[0007] Furthermore, in multi-skill collaborative task scenarios, existing technologies lack a clear skill combination planning mechanism, and the skill invocation order mainly relies on implicit reasoning of the language model, which can easily lead to problems such as unreasonable invocation order, missing prerequisite skills, or repeated invocation.

[0008] Furthermore, during the skill loading process, a large amount of skill hints and tool descriptions are simultaneously input into the language model context, leading to increased context length, higher resource consumption, and irrelevant information interfering with model inference.

[0009] Finally, existing technologies lack mechanisms for evaluating and optimizing the effectiveness of skill execution, and cannot dynamically optimize scheduling strategies based on the results of skill invocation, thus affecting the long-term performance of the system.

[0010] Therefore, how to construct a structured skill scheduling mechanism, reduce the scale of the decision-making space, improve the accuracy and stability of skill selection, and achieve reasonable scheduling and continuous optimization of multi-skill collaborative tasks in the face of a constantly increasing number of skills has become a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0011] The purpose of this invention is to overcome the shortcomings of the prior art and propose a multi-skill agent scheduling method based on skill ability relationship graphs, so as to realize structured modeling of the skill scheduling process, reduce the decision space of skill selection, improve the accuracy and stability of skill selection, and support the rational scheduling and continuous optimization of multi-skill collaborative tasks.

[0012] To achieve the above objectives, the technical solution provided by this invention is: a multi-skill agent scheduling method based on a skill ability relationship graph, comprising the following steps:

[0013] S1: Perform structured modeling of skills in a language model-driven multi-skill intelligent agent system, and construct a skill capability relationship graph. Each skill node in the skill capability relationship graph includes skill description information, capability tag information represented in the form of a set of capability tags, and skill execution attribute information including call complexity index and historical execution success rate. The skill node association relationship is established based on the dependency relationship, mutual exclusion relationship and hierarchical relationship between skills.

[0014] S2: The multi-skill intelligent agent system acquires user task requests, performs semantic parsing on user task requests, and extracts a set of ability tags related to the task so as to match them with the skill ability relationship graph constructed in step S1.

[0015] S3: For each skill node in the skill ability relationship graph, calculate the semantic relevance of each skill node based on the degree of matching between the ability tag set of each skill node and the ability tag set obtained in step S2;

[0016] S4: Based on the semantic relevance of each skill node obtained in step S3, and combined with the historical execution success rate of skills and the complexity index of skill invocation, calculate the comprehensive score of each skill node.

[0017] S5: Based on the comprehensive score of each skill node obtained in step S4, the skill nodes in the skill ability relationship graph are filtered to retain high-scoring skill nodes and remove low-scoring skill nodes, thereby obtaining a set of candidate skills to narrow down the range of skill selection.

[0018] S6: Based on the dependency, mutual exclusion and hierarchical relationships in the skill ability relationship graph, perform path constraints and path search on the candidate skill set obtained in step S5 to generate several candidate skill execution paths, and calculate the path score value of each candidate skill execution path according to the pre-designed path scoring function, thereby determining the target skill execution path;

[0019] S7: Based on the target skill execution path determined in step S6, a progressive skill disclosure mechanism is used to provide skill node information to the language model in order to guide the language model to complete multi-skill collaborative task processing;

[0020] S8: Record the number of times each skill is called, the success rate of execution, and the contribution to the task during the execution process, and update the historical execution success rate of the skill based on the records, so as to iteratively optimize the calculation of the comprehensive skill score.

[0021] Furthermore, in step S1, the set of ability tags is used to describe the type of task ability possessed by the skill, and it consists of multiple ability tags, each of which represents a task ability.

[0022] Furthermore, in step S3, the semantic relevance is calculated according to the following formula:

[0023] ;

[0024] In the formula: This represents the semantic relevance of the i-th skill node; This represents the set of ability tags for the i-th skill node; This represents the set of capability tags obtained in step S2; , indicating the skill node number; This represents the total number of skill nodes in the skill ability relationship graph.

[0025] Furthermore, in step S4, the comprehensive score is calculated according to the following formula:

[0026] ;

[0027] In the formula: This represents the overall score value of the i-th skill node; This represents the semantic relevance of the i-th skill node; This represents the historical execution success rate of the i-th skill node; This represents the complexity index of calling the i-th skill node; , , These are adjustable weighting coefficients.

[0028] Furthermore, in step S5, skill nodes are filtered based on the comprehensive score, including one or a combination of the following methods:

[0029] Method 1: Sort the skills by comprehensive score from highest to lowest and select the top K skill nodes to generate a candidate skill set;

[0030] Method 2: Generate a candidate skill set by filtering skill nodes that meet the scoring thresholds;

[0031] Wherein, K represents the preset number of candidate skills. The value of K is preset based on the total number of skill nodes in the skill ability relationship graph. When multiple skill nodes have the same comprehensive score, the skill nodes with higher historical execution success rates are given priority to be included in the candidate skill set.

[0032] Furthermore, in step S6, the path constraint includes:

[0033] Dependency constraint: If a candidate skill has a prerequisite dependent skill and that prerequisite dependent skill is not included in the candidate skill set, then the prerequisite dependent skill will be automatically added to the path, and the path execution order will follow the priority of prerequisite dependencies.

[0034] Mutually exclusive constraints: mutually exclusive skills cannot appear on the same path at the same time; only mutually exclusive skills with higher overall scores are retained on the same path.

[0035] Hierarchical relationship constraints: Upper-level skills are composite functional skills, and lower-level skills are atomic functional skills derived from the upper-level skills. Using a lower-level skill alone does not require loading the upper-level skill, while using an upper-level skill automatically includes its lower-level skill. The same path does not contain the upper-level skill and its corresponding lower-level skill repeatedly.

[0036] Furthermore, in step S6, the path score of the candidate skill execution path is calculated according to the following formula:

[0037] ;

[0038] In the formula: Indicates the first The path score of each candidate skill execution path; This represents the overall score of the i-th skill node in the path; Indicates the first The number of skill nodes in each candidate skill execution path;

[0039] The path score is obtained by summing the comprehensive score of each skill node in the path and normalizing it in combination with the path length, thereby comprehensively reflecting the overall capability matching degree and execution complexity of the skill execution path.

[0040] Furthermore, in step S7, the progressive skill disclosure mechanism is specifically as follows:

[0041] In the initial stage, the language model is provided with the information of the starting skill node in the target skill execution path; after the language model calls the current skill node and obtains the execution result, the information of the subsequent skill nodes in the target skill execution path is gradually opened up based on the skill ability relationship graph; by gradually expanding the range of optional skills of the language model, the language model is guided to complete the multi-skill collaborative task processing according to the target skill execution path.

[0042] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0043] 1. By constructing a skill ability relationship graph, skills are structurally modeled, and dependency, mutual exclusion and hierarchical relationships are introduced to realize the explicit expression of the relationship between skills, thereby improving the rationality and interpretability of skill scheduling.

[0044] 2. By using a semantic relevance calculation method based on a set of capability tags, a precise match between user task requirements and skill capabilities is achieved, thereby improving the accuracy of skill selection.

[0045] 3. By introducing historical execution success rate and call complexity indicators to construct a comprehensive score, a multi-dimensional skill assessment is achieved, thereby improving the scientific nature and stability of skill selection.

[0046] 4. By filtering skill nodes to generate a candidate skill set, the range of skill selection is effectively narrowed, thereby reducing the decision space of the language model and improving reasoning efficiency.

[0047] 5. By performing path search based on the skill ability relationship graph and constructing candidate skill execution paths, structured scheduling of multi-skill collaborative tasks is realized, effectively avoiding the problems of unreasonable skill call order and repeated calls.

[0048] 6. The path scoring function is used to evaluate the execution paths of candidate skills and select the optimal path, which further improves the execution effect of multi-skill combination tasks.

[0049] 7. By using a progressive skill disclosure mechanism, skill information is provided to the language model on demand, effectively controlling the model context size, reducing token consumption, and minimizing interference from irrelevant information.

[0050] 8. By recording the number of times skills are invoked, the success rate of execution, and the contribution to the task during the skill execution process, and dynamically updating the historical success rate of skill execution, the ability to continuously optimize the skill scheduling strategy is realized.

[0051] 9. This invention improves the scheduling efficiency, execution stability, and system scalability of multi-skill intelligent agent systems in complex task scenarios. Attached Figure Description

[0052] Figure 1 This is a schematic diagram of the overall process of the method of the present invention.

[0053] Figure 2 This is a schematic diagram of the structure of a skill-ability relationship graph.

[0054] Figure 3 This is a schematic diagram of the comprehensive score calculation process.

[0055] Figure 4 This is a schematic diagram illustrating the generation and scoring of candidate skill execution paths.

[0056] Figure 5 This is a schematic diagram illustrating the implementation of a progressive skills disclosure mechanism. Detailed Implementation

[0057] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments, but the implementation of the present invention is not limited thereto.

[0058] like Figure 1 As shown in the figure, this embodiment discloses a multi-skill agent scheduling method based on skill ability relationship graph, which is applicable to multi-skill agent systems driven by large language models. The specific details are as follows:

[0059] S1, Construct a skill and ability relationship graph:

[0060] First, a structured model of multiple skills in the multi-skill intelligent agent system is performed, and a skill capability relationship graph is constructed. Each skill node includes the following information:

[0061] Skill description information, which describes the function and scope of application of the skill;

[0062] A set of ability tags used to describe the types of task abilities that the skill possesses;

[0063] Skill execution attribute information, including invocation complexity metrics and historical execution success rate .

[0064] Among them, the call complexity metric The value ranges from [0,1] and is used to characterize the resource consumption, execution latency, or tool invocation complexity required to invoke the skill. A higher value indicates higher invocation complexity. Historical execution success rate. The value ranges from [0,1] and is used to characterize the success rate of the skill in historical tasks. The larger the value, the higher the historical success rate.

[0065] like Figure 2 As shown, this embodiment constructs the following skill nodes:

[0066]

[0067] The following relationships are established between the skill nodes:

[0068] Dependency relationships: S2 depends on S1; S3 depends on S2.

[0069] Mutual exclusion relationship: S6 and S7 are mutually exclusive.

[0070] Hierarchical relationship: S5 is the upper-level skill, which is a composite skill; S1, S2, S3, and S4 are the lower-level skills of S5, which are atomic skills derived from the upper-level skills.

[0071] The skill and ability relationship map is formed in the above way.

[0072] S2, parsing user task requests:

[0073] The intelligent agent system obtains user task requests, performs semantic parsing, and extracts a set of task capability tags.

[0074] In this embodiment, the user input request is: "Please identify the disease in the rice leaf image, analyze the cause of the disease, provide a prevention and control plan, and formulate comprehensive management suggestions based on the weather conditions for the next three days."

[0075] The system analyzes and obtains the set of task capability tags:

[0076] T = {Image recognition, disease identification, disease analysis, prevention and control plan, weather forecast, integrated management}

[0077] The number of elements in the set is:

[0078] ;

[0079] S3, Calculate the semantic relevance of skill nodes:

[0080] For each skill node in the skill ability relationship graph, the semantic relevance of each skill node is calculated based on the degree of matching between the set of ability labels of the skill node and the set of task ability labels obtained in step S2.

[0081] Semantic relevance is calculated using the following formula:

[0082] ;

[0083] In the formula: This represents the semantic relevance of the i-th skill node; This represents the set of ability tags for the i-th skill node; This represents the set of capability tags obtained in step S2; , indicating the skill node number; This represents the total number of skill nodes in the skill ability relationship graph.

[0084] In this embodiment, the calculation results are as follows:

[0085]

[0086] S4, Calculate the overall skill score:

[0087] Based on the semantic relevance obtained in step S3, and combined with the historical execution success rate and call complexity index, the comprehensive score of each skill node is calculated.

[0088] The overall skill score is calculated using the following formula:

[0089] ;

[0090] In the formula: This represents the overall score value of the i-th skill node; This represents the semantic relevance of the i-th skill node; This represents the historical execution success rate of the i-th skill node; This represents the complexity index of calling the i-th skill node; , , These are adjustable weighting coefficients.

[0091] In this embodiment, we take:

[0092] ;

[0093] The calculation results are as follows:

[0094]

[0095] like Figure 3 As shown, the above method enables multi-dimensional quantitative evaluation of skill nodes.

[0096] S5, Generate a set of candidate skills:

[0097] Based on the comprehensive score obtained in step S4, the skill nodes are filtered to retain high-scoring skill nodes and remove low-scoring skill nodes, thereby obtaining a set of candidate skills.

[0098] This embodiment uses a method of selecting the top K skill nodes after sorting by their rating values, where:

[0099] K=5;

[0100] Sort by rating from highest to lowest:

[0101] S1, S4, S3, S7, S2, S5, S6;

[0102] Obtain a set of candidate skills:

[0103] C = {S1, S4, S3, S7, S2};

[0104] Since S3 depends on S2, and S2 is already in the set, the dependency constraint is satisfied.

[0105] S6, Determine the target skill execution path:

[0106] Based on the skill node relationships in the skill ability relationship graph, a path search is performed on the candidate skill set to generate several candidate skill execution paths.

[0107] like Figure 4 As shown, this embodiment generates the following candidate paths:

[0108] Path P1: S1→S2→S3;

[0109] Path P2: S1→S2→S3→S4;

[0110] Path P3: S1→S2→S7→S4;

[0111] Path P4:S5;

[0112] in:

[0113] P4 calls upon the upper-level skill S5, automatically covering the lower-level skill.

[0114] The path score is calculated using the following formula:

[0115] ;

[0116] In the formula: Indicates the first The path score of each candidate skill execution path; This represents the overall score of the i-th skill node in the path; Indicates the first The number of skill nodes in each candidate skill execution path.

[0117] The calculation results are as follows:

[0118]

[0119] P2 and P3 have the highest and consistent path scores. When the path scores are the same, the candidate skill execution path with the higher total historical skill execution success rate can be selected as the target skill execution path. In this embodiment, P2 is selected as the target skill execution path.

[0120] This path fully covers the four requirements of identification, diagnosis, prevention and control, and weather, demonstrating the structured scheduling capability of this invention.

[0121] S7 implements a progressive skills disclosure mechanism:

[0122] Based on the target skill execution path determined in step S6, a progressive skill disclosure mechanism is used to provide skill node information to the language model.

[0123] like Figure 5 As shown, the specific process is as follows:

[0124] Phase 1

[0125] Open S1: Pest and disease image recognition skills.

[0126] Output: "The result is rice sheath blight".

[0127] Phase Two

[0128] Open S2: Disease diagnosis skills.

[0129] Output: "Currently in the initial stage, high humidity in the field increases the risk of transmission."

[0130] Phase Three

[0131] Unlock S3: Recommended skills for prevention and control.

[0132] Output: "It is recommended to use Jinggangmycin-type agents for early prevention and treatment."

[0133] Phase 4

[0134] Unlock S4: Weather query skill.

[0135] Output: "Continuous rain is expected for the next three days. It is recommended to apply pesticides during breaks in the rain and to improve drainage."

[0136] The final language model generates a comprehensive response.

[0137] By gradually releasing skill node information, irrelevant skill information can be reduced from entering the context, thereby reducing token consumption and improving inference stability.

[0138] S8, Update skill execution history success rate:

[0139] After the task is completed, the number of times each skill is called, the success rate of execution, and the contribution to the task are recorded, and the historical execution success rate of the skill is updated based on the records.

[0140] In this embodiment, the task was successfully executed, and the following update method was adopted:

[0141] ;

[0142] In the formula: This indicates the updated historical execution success rate; This indicates the historical execution success rate before the update; This indicates the result of the execution, with 1 for success and 0 for failure. The smoothing coefficient is set to 0.9.

[0143] This task was successfully completed, therefore:

[0144] ;

[0145] The calculation results are as follows:

[0146]

[0147] The updated historical execution success rate will be included in the calculation of the overall score to achieve continuous optimization of the scheduling strategy.

[0148] Implementation Results: This invention achieves structured scheduling of multi-skill intelligent agent systems through skill ability relationship graphs, scoring and screening, path planning, and a progressive skill disclosure mechanism.

[0149] Compared to existing methods that rely solely on implicit reasoning using language models, this invention offers the following advantages:

[0150] 1. Reduce the decision-making space for skill selection;

[0151] 2. Improve the accuracy of skill matching;

[0152] 3. Optimize the order of multi-skill activation;

[0153] 4. Reduce context token consumption;

[0154] 5. Improve system stability and scalability;

[0155] 6. Possesses long-term self-optimization capabilities.

[0156] The above embodiments are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any equivalent substitutions, process adjustments, parameter changes, or structural modifications made based on the content of this specification shall fall within the scope of protection of the present invention.

Claims

1. A multi-skill agent scheduling method based on a skill ability relationship graph, characterized in that, Includes the following steps: S1: Perform structured modeling of skills in a language model-driven multi-skill intelligent agent system, and construct a skill capability relationship graph. Each skill node in the skill capability relationship graph includes skill description information, capability tag information represented in the form of a set of capability tags, and skill execution attribute information including call complexity index and historical execution success rate. The skill node association relationship is established based on the dependency relationship, mutual exclusion relationship and hierarchical relationship between skills. S2: The multi-skill intelligent agent system acquires user task requests, performs semantic parsing on user task requests, and extracts a set of ability tags related to the task so as to match them with the skill ability relationship graph constructed in step S1. S3: For each skill node in the skill ability relationship graph, calculate the semantic relevance of each skill node based on the degree of matching between the ability tag set of each skill node and the ability tag set obtained in step S2; S4: Based on the semantic relevance of each skill node obtained in step S3, and combined with the historical execution success rate of skills and the complexity index of skill invocation, calculate the comprehensive score of each skill node. S5: Based on the comprehensive score of each skill node obtained in step S4, the skill nodes in the skill ability relationship graph are filtered to retain high-scoring skill nodes and remove low-scoring skill nodes, thereby obtaining a set of candidate skills to narrow down the range of skill selection. S6: Based on the dependency, mutual exclusion and hierarchical relationships in the skill ability relationship graph, perform path constraints and path search on the candidate skill set obtained in step S5 to generate several candidate skill execution paths, and calculate the path score value of each candidate skill execution path according to the pre-designed path scoring function, thereby determining the target skill execution path; S7: Based on the target skill execution path determined in step S6, a progressive skill disclosure mechanism is used to provide skill node information to the language model in order to guide the language model to complete multi-skill collaborative task processing; S8: Record the number of times each skill is called, the success rate of execution, and the contribution to the task during the execution process, and update the historical execution success rate of the skill based on the records, so as to iteratively optimize the calculation of the comprehensive skill score.

2. The multi-skill agent scheduling method based on skill ability relationship graph as described in claim 1, characterized in that, In step S1, the set of ability tags is used to describe the types of task abilities possessed by the skill. It consists of multiple ability tags, each of which represents a task ability.

3. The multi-skill agent scheduling method based on skill ability relationship graphs according to claim 1, characterized in that, In step S3, the semantic relevance is calculated according to the following formula: ； In the formula: This represents the semantic relevance of the i-th skill node; This represents the set of ability tags for the i-th skill node; This represents the set of capability tags obtained in step S2; , indicating the skill node number; This represents the total number of skill nodes in the skill ability relationship graph.

4. The multi-skill agent scheduling method based on skill ability relationship graph as described in claim 1, characterized in that, In step S4, the comprehensive score is calculated according to the following formula: ； In the formula: This represents the overall score value of the i-th skill node; This represents the semantic relevance of the i-th skill node; This represents the historical execution success rate of the i-th skill node; This represents the complexity index of calling the i-th skill node; , , These are adjustable weighting coefficients.

5. A multi-skill agent scheduling method based on a skill ability relationship graph according to claim 1, characterized in that, In step S5, skill nodes are filtered based on the comprehensive score, including one or a combination of the following methods: Method 1: Sort the skills by comprehensive score from highest to lowest and select the top K skill nodes to generate a candidate skill set; Method 2: Generate a candidate skill set by filtering skill nodes that meet the scoring thresholds; Wherein, K represents the preset number of candidate skills. The value of K is preset based on the total number of skill nodes in the skill ability relationship graph. When multiple skill nodes have the same comprehensive score, the skill nodes with higher historical execution success rates are given priority to be included in the candidate skill set.

6. The multi-skill agent scheduling method based on skill ability relationship graph as described in claim 1, characterized in that, In step S6, the path constraints include: Dependency constraint: If a candidate skill has a prerequisite dependent skill and that prerequisite dependent skill is not included in the candidate skill set, then the prerequisite dependent skill will be automatically added to the path, and the path execution order will follow the priority of prerequisite dependencies. Mutually exclusive constraints: mutually exclusive skills cannot appear on the same path at the same time; only mutually exclusive skills with higher overall scores are retained on the same path. Hierarchical relationship constraints: Upper-level skills are composite functional skills, and lower-level skills are atomic functional skills derived from the upper-level skills. Using a lower-level skill alone does not require loading the upper-level skill, while using an upper-level skill automatically includes its lower-level skill. The same path does not contain the upper-level skill and its corresponding lower-level skill repeatedly.

7. The multi-skill agent scheduling method based on skill ability relationship graphs according to claim 1, characterized in that, In step S6, the path score of the candidate skill execution path is calculated according to the following formula: ； In the formula: Indicates the first The path score of each candidate skill execution path; This represents the overall score of the i-th skill node in the path; Indicates the first The number of skill nodes in each candidate skill execution path; The path score is obtained by summing the comprehensive score of each skill node in the path and normalizing it in combination with the path length, thereby comprehensively reflecting the overall capability matching degree and execution complexity of the skill execution path.

8. The multi-skill agent scheduling method based on skill ability relationship graph as described in claim 1, characterized in that, In step S7, the progressive skill disclosure mechanism is specifically as follows: In the initial stage, the language model is provided with the information of the starting skill node in the target skill execution path; after the language model calls the current skill node and obtains the execution result, the information of the subsequent skill nodes in the target skill execution path is gradually opened based on the skill ability relationship graph. By gradually expanding the range of selectable skills in the language model, the language model is guided to complete multi-skill collaborative task processing according to the target skill execution path.

Images