Intelligent agent skill evolution method and apparatus, storage medium, and electronic device

Abstract

This invention provides a method, apparatus, storage medium, and electronic device for evolving intelligent agent skills, relating to the field of artificial intelligence technology. The method includes: invoking a first intelligent agent to infer the task execution input data of each task according to initial intelligent agent skills, obtaining the execution trajectory and inference output of each task; decomposing the initial intelligent agent skills to obtain multiple sub-skills; determining the utility value of each sub-skill based on the execution trajectory and inference output of each task; identifying a sub-skill to be evolved from the multiple sub-skills based on the utility values ​​of each sub-skill; invoking a second intelligent agent to adjust the sub-skill to be evolved, obtaining an adjusted intelligent agent skill; and determining a target intelligent agent skill for a target task type based on the adjusted intelligent agent skill. Embodiments of this invention can improve the accuracy of intelligent agent skills.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method, apparatus, storage medium, and electronic device for the skill evolution of an intelligent agent. Background Technology

[0002] Currently, with the widespread application of large language model agents in complex tasks such as spreadsheet manipulation, mathematical reasoning, document question answering, and tool invocation, skills (i.e., agent skills) have become the core knowledge carrier driving agents to complete tasks. Agent skills typically carry standard operating procedures, domain rules, and exception handling strategies in text form. However, related technologies usually rely on manually written skills, resulting in low accuracy. Therefore, there is currently no satisfactory solution for improving the accuracy of agent skills. Summary of the Invention

[0003] In view of this, embodiments of the present invention provide a method, apparatus, storage medium, and electronic device for the evolution of intelligent agent skills, to solve the problem of low accuracy in determining intelligent agent skills in related technologies. In other words, embodiments of the present invention can adjust the sub-skills to be evolved in the initial intelligent agent skills through an intelligent agent to obtain adjusted intelligent agent skills, thereby determining the target intelligent agent skill with a higher accuracy rate. This can realize the automated evolution of intelligent agent skills, effectively improve the accuracy of the target intelligent agent skill, and the embodiments of the present invention can make the evolution direction of intelligent agent skills clear through the sub-skills to be evolved, no longer blindly modifying them, thus changing from "global random modification" to "precise positioning of the sub-skills to be evolved", making the evolution process highly controllable, thereby effectively improving the evolution efficiency.

[0004] According to one aspect of the present invention, a method for the evolution of agent skills is provided, the method comprising: Obtain the initial agent skills corresponding to the target task type, and obtain the training data set; the training data set includes the training data of each task in multiple tasks under the target task type, and each training data includes the task execution input data of one task; The first intelligent agent is invoked to reason about the task execution input data of each task according to the initial intelligent agent skills, so as to obtain the execution trajectory and reasoning output of each task; The initial agent skill is decomposed to obtain multiple sub-skills; and based on the execution trajectory and inference output of each task, the utility value of each sub-skill is determined. Based on the utility value of each sub-skill, a sub-skill to be evolved is determined from the plurality of sub-skills, the sub-skill to be evolved including the sub-skill with the smallest utility value among the plurality of sub-skills; and a second intelligent agent is invoked to adjust the sub-skill to be evolved, resulting in an adjusting intelligent agent skill; Based on the adjusted agent skills, the target agent skills for the target task type are determined.

[0005] According to another aspect of the present invention, an agent skill evolution device is provided, the device comprising: The acquisition unit is used to acquire the initial agent skills corresponding to the target task type and to acquire the training data set; the training data set includes the training data of each task in multiple tasks under the target task type, and each training data includes the task execution input data of one task; The processing unit is used to invoke the first intelligent agent to reason about the task execution input data of each task according to the initial intelligent agent skills, so as to obtain the execution trajectory and reasoning output of each task; The processing unit is further configured to decompose the initial agent skill into multiple sub-skills; and based on the execution trajectory and inference output of each task, determine the utility value of each sub-skill among the multiple sub-skills. The processing unit is further configured to determine, based on the utility value of each sub-skill, a sub-skill to be evolved from the plurality of sub-skills, the sub-skill to be evolved including the sub-skill with the smallest utility value among the plurality of sub-skills; and to invoke a second agent to adjust the sub-skill to be evolved, thereby obtaining an adjusting agent skill; The processing unit is further configured to determine the target agent skill of the target task type based on the adjusted agent skill.

[0006] According to another aspect of the present invention, an electronic device is provided, the electronic device including a processor and a memory storing a program, wherein the program includes instructions that, when executed by the processor, cause the processor to perform the methods mentioned above.

[0007] According to another aspect of the present invention, a non-transitory computer-readable storage medium is provided storing computer instructions for causing a computer to perform the methods mentioned above.

[0008] This invention provides an embodiment that can acquire initial agent skills corresponding to a target task type, as well as a training data set. The training data set includes training data for each task among multiple tasks under the target task type, with each training data set including the task execution input data for one task. A first agent is invoked to infer the task execution input data for each task according to the initial agent skills, obtaining the execution trajectory and inference output for each task. Then, the initial agent skills can be decomposed to obtain multiple sub-skills. Based on the execution trajectory and inference output of each task, the utility value of each sub-skill is determined. Further, based on the utility values ​​of each sub-skill, a sub-skill to be evolved can be determined from the multiple sub-skills. A second agent is invoked to adjust the sub-skill to be evolved, obtaining an adjusted agent skill. Correspondingly, the target agent skill for the target task type can be determined based on the adjusted agent skill. As can be seen, the embodiments of the present invention can adjust the sub-skills to be evolved in the initial intelligent agent skills through the intelligent agent, obtain the adjusted intelligent agent skills, and then determine the target intelligent agent skills with higher accuracy, thereby realizing the automated evolution of intelligent agent skills and effectively improving the accuracy of target intelligent agent skills. Furthermore, the embodiments of the present invention can make the evolution direction of intelligent agent skills clear through the sub-skills to be evolved, no longer blindly modifying them, thus changing from "global random modification" to "precise positioning of the sub-skills to be evolved", making the evolution process highly controllable, thereby effectively improving the evolution efficiency. Attached Figure Description

[0009] Further details, features, and advantages of the invention are disclosed in the following description of exemplary embodiments in conjunction with the accompanying drawings, in which: Figure 1 A flowchart illustrating an agent skill evolution method according to an exemplary embodiment of the present invention is shown; Figure 2 A flowchart illustrating another agent skill evolution method according to an exemplary embodiment of the present invention is shown; Figure 3 A flowchart illustrating yet another method for the evolution of agent skills according to an exemplary embodiment of the present invention is shown; Figure 4 A schematic block diagram of an agent skill evolution device according to an exemplary embodiment of the present invention is shown; Figure 5 A structural block diagram of an exemplary electronic device that can be used to implement embodiments of the present invention is shown. Detailed Implementation

[0010] Embodiments of the present invention will now be described in more detail with reference to the accompanying drawings. While some embodiments of the invention are shown in the drawings, it should be understood that the invention can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of the invention. It should be understood that the accompanying drawings and embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the invention.

[0011] It should be understood that the various steps described in the method embodiments of the present invention may be performed in different orders and / or in parallel. Furthermore, the method embodiments may include additional steps and / or omit the steps shown. The scope of the present invention is not limited in this respect.

[0012] The term "comprising" and its variations as used herein are open-ended, meaning "including but not limited to". The term "based on" means "at least partially based on". The term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one additional embodiment"; the term "some embodiments" means "at least some embodiments". Definitions of other terms will be given in the following description. It should be noted that the concepts of "first", "second", etc., mentioned in this invention are used only to distinguish different devices, modules, or units, and are not intended to limit the order of functions performed by these devices, modules, or units or their interdependencies.

[0013] It should be noted that the terms "a" and "a plurality of" used in this invention are illustrative rather than restrictive. Those skilled in the art should understand that, unless otherwise expressly indicated in the context, they should be understood as "one or more".

[0014] The names of the messages or information exchanged between the multiple devices in the embodiments of the present invention are for illustrative purposes only and are not intended to limit the scope of these messages or information.

[0015] It should be noted that the executing entity of the intelligent agent skill evolution method provided in this embodiment of the invention can be one or more electronic devices, and this invention does not limit this; wherein, the electronic device can be a terminal (i.e., a client) or a server. Therefore, when the executing entity includes multiple electronic devices, and among these multiple electronic devices are at least one terminal and at least one server, the intelligent agent skill evolution method provided in this embodiment of the invention can be jointly executed by the terminal and the server. Accordingly, the terminal mentioned herein can include, but is not limited to: smartphones, tablets, laptops, desktop computers, smartwatches, smart voice interaction devices, smart home appliances, vehicle terminals, aircraft, etc. The server mentioned herein can be an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN (Content Delivery Network), and big data and artificial intelligence platforms, etc.

[0016] Based on the above description, this embodiment of the invention proposes a method for the evolution of intelligent agent skills. This method can be executed by the aforementioned electronic device (terminal or server); or, it can be executed jointly by a terminal and a server. For ease of explanation, the following description will use the execution of this method by an electronic device as an example; such as... Figure 1 As shown, the method for evolving the skills of an intelligent agent may include the following steps S101-S105: S101, obtain the initial agent skills corresponding to the target task type, and obtain the training data set; the training data set includes the training data of each task in multiple tasks under the target task type, and each training data includes the task execution input data of one task.

[0017] Optionally, the target task type can be any task type, and this embodiment of the invention does not limit this; for example, the target task type can be a Chinese-to-English translation task type, or a flight scheduling and conflict handling task type (which can also be simply referred to as a flight scheduling task type, etc.), and so on. Optionally, the number of target task types can be one or more, and this embodiment of the invention does not limit this.

[0018] Optionally, the aforementioned initial agent skill (also referred to as initial skill) can be manually programmed by the user. For example, the agent skill evolution operation performed by the user can carry the initial agent skill corresponding to the target task type. In this case, the initial agent skill corresponding to the target task type can be obtained from the data carried by the agent skill evolution operation. Alternatively, the initial agent skill can be automatically generated by an agent (such as any agent). For example, the agent can generate the initial agent skill through the task information of the target task type to obtain the initial agent skill corresponding to the target task type. Alternatively, the electronic device stores the initial agent skill corresponding to the target task type in its own storage space. In this case, the electronic device can obtain the initial agent skill corresponding to the target task type from its own storage space, and so on. This embodiment of the invention does not limit this.

[0019] Optionally, the training dataset can be obtained in the following ways, including but not limited to: The first acquisition method: The electronic device has a set of training data for the target task type stored in its own storage space. In this case, the electronic device can obtain the set of training data from its own storage space.

[0020] The second method is to use electronic devices to obtain training dataset download links and download the training dataset using those links.

[0021] The third acquisition method: When an electronic device detects an upload operation of a training data set for a target task type, it can use the data set indicated by the upload operation as the training data set to acquire the training data set, and so on.

[0022] Optionally, the training dataset may include user-generated test data (i.e., task execution input data) and / or corresponding outputs (i.e., result labels, also known as output labels), etc.; this embodiment of the invention does not limit this. Optionally, the result labels of a task may be set according to experience or actual needs; this embodiment of the invention does not limit this.

[0023] Optionally, the electronic device can also acquire task information of the target task type. This task information can be used to indicate the problem that needs to be solved for the target task type, i.e., it can be used to indicate the problem that needs to be solved for a task belonging to the target task type. It should be noted that the specific content of the task information is not limited in the embodiments of the present invention. That is, the task information can be set according to experience or actual needs, and the embodiments of the present invention do not limit it in this regard. For example, when the target task type is a Chinese-to-English translation task, the task information of the target task type can be: translating everyday Chinese short sentences into standard English, etc. Optionally, the electronic device can store the task information of the target task type in its own storage space. In this case, the electronic device can acquire the task information of the target task type from its own storage space; or, when a task information setting operation for the target task type is detected, the task information indicated by the task information setting operation can be used as the task information of the target task type to achieve the acquisition of the task information of the target task type, etc. The embodiments of the present invention do not limit it in this regard.

[0024] Optionally, any two of the above-mentioned intelligent agent skill evolution operation, training data set upload operation, and task information setting operation may be executed simultaneously or not simultaneously, and the embodiments of the present invention do not limit this; for example, the intelligent agent skill evolution operation may include the training data set upload operation and the task information setting operation, etc.

[0025] S102, invoke the first intelligent agent, and according to the initial intelligent agent skills, reason about the task execution input data of each task to obtain the execution trajectory and reasoning output of each task.

[0026] Optionally, the first intelligent agent can be any intelligent agent (such as including but not limited to any large language model, etc.), and the embodiments of the present invention do not limit this; that is, the embodiments of the present invention do not limit the specific structure of the first intelligent agent.

[0027] Based on this, the electronic device can drive the first intelligent agent according to the initial intelligent agent skills to infer the task execution input data of any task, so as to realize the execution of any task by driving the first intelligent agent according to the initial intelligent agent skills, and obtain the execution trajectory (also known as behavior trajectory or trajectory, etc.) and inference output (also known as inference result, etc.) of any task. Accordingly, the embodiments of the present invention can realize the execution of various tasks by driving the first intelligent agent through the initial intelligent agent skills, thereby collecting the execution process and inference output of each trajectory (i.e., the trajectory of each task).

[0028] For example, an electronic device can input task execution input data for any task (e.g., in the case of a Chinese-to-English translation task, task execution input data can be a piece of Chinese or a Chinese text) and / or task information of the target task type into a first intelligent agent, so as to drive the first intelligent agent to execute any task according to the initial intelligent agent skills, thereby outputting the execution trajectory and inference output of any task through the first intelligent agent, and so on.

[0029] S103, decompose the initial agent skill into multiple sub-skills; and determine the utility value of each sub-skill based on the execution trajectory and inference output of each task.

[0030] Optionally, when decomposing (or segmenting) the initial agent skill to obtain multiple sub-skills, the electronic device can determine a skill decomposition strategy corresponding to the target task type. This skill decomposition strategy can be a delimiter segmentation strategy or a semantic segmentation strategy. The electronic device can then decompose the initial agent skill according to this strategy to obtain multiple sub-skills. In this embodiment of the invention, a sub-skill can also be referred to as a skill fragment.

[0031] Optionally, one task type can correspond to one skill decomposition strategy; optionally, the skill decomposition strategies corresponding to different task types can be the same or different, and this embodiment of the invention does not limit this. Optionally, the skill decomposition strategy corresponding to a task type can be set according to experience or actual needs, and this embodiment of the invention does not limit this.

[0032] Optionally, the above-mentioned delimiter decomposition strategy may instruct skill decomposition (which can also be represented as segmentation) according to at least one delimiter; optionally, the at least one delimiter may include, but is not limited to, at least one of the following: comma, period, semicolon, and question mark, etc., and the embodiments of the present invention do not limit this. Optionally, the above-mentioned semantic segmentation strategy may instruct skill decomposition according to semantic integrity, so as to decompose the agent's skills into non-overlapping sub-skills that support independent attribution and independent modification. Based on this, multiple sub-skills can be non-overlapping sub-skills that support independent attribution and independent modification.

[0033] Optionally, when decomposing the initial agent skill according to the skill decomposition strategy to obtain multiple sub-skills of the initial agent skill, if the skill decomposition strategy is a separator segmentation strategy, the initial agent skill can be decomposed according to at least one separator to obtain multiple sub-skills of the initial agent skill. That is, the separator position in the initial agent skill can be determined according to at least one separator (i.e., the character is the position of any separator in at least one separator), and the separator position in the initial agent skill can be segmented to achieve skill decomposition of the initial agent skill. If the skill decomposition strategy is a semantic segmentation strategy, a semantic segmentation agent can be called to decompose the initial agent skill to obtain multiple sub-skills of the initial agent skill. For example, the initial agent skill can be input into the semantic segmentation agent to output multiple sub-skills, etc. Optionally, the number of separator positions in an agent skill can be one or more, and this embodiment of the invention does not limit this. Optionally, the semantic segmentation agent can be any agent, that is, this embodiment of the invention does not limit the specific structure of the semantic segmentation agent. Optionally, when decomposing skills through a semantic segmentation agent, semantic segmentation prompts can be input, such as semantic segmentation of skills; optionally, semantic segmentation prompts can be set according to experience or actual needs, and this embodiment of the invention does not limit this.

[0034] Optionally, in other embodiments, the electronic device may also directly decompose the initial agent skill according to at least one delimiter, in which case agent skills of different task types are all decomposed according to at least one delimiter; or, the initial agent skill may be directly decomposed through a semantic segmentation agent, in which case agent skills of different task types can be decomposed through a semantic segmentation agent, etc.; the present invention does not limit this.

[0035] S104, based on the utility value of each sub-skill, determine the sub-skill to be evolved from multiple sub-skills; and call the second agent to adjust the sub-skill to be evolved, thus obtaining the adjusted agent skill.

[0036] In one implementation, when determining the sub-skill to be evolved from multiple sub-skills based on the utility value of each sub-skill, the sub-skill with the smallest utility value can be selected and used as the sub-skill to be evolved. The sub-skill with the smallest utility value can also be called the weakest utility sub-skill.

[0037] In another implementation, the electronic device can also determine the evolution ratio and calculate the number of sub-skills to be evolved using the evolution ratio and the total number of sub-skills among multiple sub-skills. Correspondingly, based on the utility value of each sub-skill, the top 10 sub-skills with the smallest utility value can be determined from among the multiple sub-skills, and these determined sub-skills can be used as the sub-skills to be evolved. In this case, the number of sub-skills to be evolved can be the number of sub-skills to be evolved, and so on. Optionally, when calculating the number of sub-skills to be evolved, the product between the evolution ratio and the total number of sub-skills among multiple sub-skills can be rounded up to obtain the number of sub-skills to be evolved; or, the product between the evolution ratio and the total number of sub-skills among multiple sub-skills can be rounded down, and when the rounding down results in 0, the number of sub-skills to be evolved can be updated to 1, and so on. This embodiment of the invention does not limit this. Optionally, the evolution ratio can be set according to experience or actual needs, and this embodiment of the invention does not limit this.

[0038] Based on this, the sub-skills to be evolved may include, but are not limited to, the sub-skill with the lowest utility value among multiple sub-skills, and so on.

[0039] Optionally, the second intelligent agent can be any intelligent agent, and the embodiments of the present invention do not limit this; that is, the embodiments of the present invention do not limit the specific structure of the second intelligent agent.

[0040] S105, Based on adjusting the agent's skills, determine the target agent's skills for the target task type.

[0041] Optionally, the electronic device can use the adjusted agent skill as the initial agent skill (i.e., as the initial agent skill for the next iteration) and iteratively execute the above-mentioned call to the first agent. According to the initial agent skill, it performs reasoning on the task execution input data of each task to achieve continuous evolution of the agent skill until the evolution stopping condition is met. The evolution stopping condition is used to determine whether to terminate the evolution iteration. That is, the next iteration can be used as the current iteration to continue to adjust the sub-skills to be evolved in the initial agent under the current iteration to obtain the adjusted agent skill under the current iteration until the evolution stopping condition (i.e., the iteration stopping condition) is met.

[0042] Based on this, electronic devices can use the adjusting agent skill when the evolutionary stopping condition is met (i.e., the adjusting agent skill in the current iteration) as the target agent skill for the target task type. The process of determining whether the evolutionary stopping condition is met can also be called convergence judgment; that is, convergence is determined when the evolutionary stopping condition is met, thus outputting the final skill (i.e., the target agent skill), such as... Figure 2As shown. In this embodiment of the invention, determining the contribution of each sub-skill to the task (such as determining the contribution value of any sub-skill to each task) can be a specific implementation of trajectory-skill fragment attribution, and determining the utility value of each sub-skill can also be expressed as a skill fragment utility calculation process.

[0043] Optionally, the evolutionary stopping condition can be set according to experience or actual needs, and this embodiment of the invention does not limit this. For example, the evolutionary stopping condition may include, but is not limited to, at least one of the following: the number of iterations reaches the maximum number of iterations (also known as a preset iteration threshold), the accuracy (i.e., the accuracy of the current skill on the training dataset, also known as the accuracy of the current skill, which may be the skill of the adjusted agent) reaches a preset accuracy threshold, etc., and this embodiment of the invention does not limit this. Based on this, when any one of the evolutionary stopping conditions is met, it can be determined that the evolutionary stopping condition is met, such as the number of iterations reaching the maximum number of iterations or the accuracy reaching a preset accuracy threshold, etc. When each of the evolutionary stopping conditions is not met, it can be determined that the evolutionary stopping condition is not met, such as the number of iterations not reaching the maximum number of iterations and the accuracy not reaching the preset accuracy threshold, etc.; for example, the evolutionary stopping condition can be expressed as: (Acc≥A) (Iter≥K), where Acc can represent accuracy, A can represent a preset accuracy threshold, Iter can represent the number of iterations, and K can represent the maximum number of iterations. Accordingly, if any one of the evolutionary stopping conditions is met, the iteration stops; otherwise, the iteration count is incremented by 1, and iteration continues. Optionally, both the maximum number of iterations and the preset accuracy threshold can be set based on experience or actual needs, and this embodiment of the invention does not limit this. Based on this, when the evolutionary stopping conditions include the number of iterations reaching the maximum number of iterations and the accuracy reaching the preset accuracy threshold, a dual constraint of the accuracy threshold and the maximum number of iterations can be achieved. Furthermore, the convergence conditions of this embodiment of the invention are clear and can be directly embedded into an automated training platform.

[0044] Optionally, when determining accuracy, a first intelligent agent can be invoked to infer the task execution input data of each task according to the adjusted agent's skills, obtaining the execution trajectory and inference output of each task (i.e., the current execution trajectory and inference output). Based on the inference output of each task (i.e., the current inference output), the number of tasks with accurate inference output can be determined from multiple tasks, and the ratio between the number of tasks with accurate inference output and the total number of tasks in multiple tasks can be used as the accuracy rate, and so on. Optionally, when judging whether the inference output of any task is accurate, it can be judged whether the inference output of any task is consistent with the result label, that is, whether the inference output of any task conforms to the result label of any task. When the inference output of any task conforms to the result label of any task (i.e., the inference output of any task and the result label are consistent), the inference output of any task can be determined to be accurate, or it can be represented as a success of any task; when the inference output of any task does not conform to the result label of any task (i.e., the inference output of any task and the result label are inconsistent), the inference output of any task can be determined to be inaccurate (i.e., an error), or it can be represented as a failure of any task, and so on. Optionally, when determining whether the inference output of any task matches the result label of any task, it can be determined whether the inference output of any task is the same as the result label of any task. If they are the same, it can be determined that they match; if they are different, it can be determined that they do not match. Alternatively, the inference output and result label of any task can be input into the accurate judgment agent, so that the accurate judgment agent outputs a conclusion on whether the inference output of any task matches the result label or outputs a conclusion on whether the inference output of any task is accurate, etc. The embodiments of the present invention do not limit this. Optionally, the accurate judgment agent can be any agent, and the embodiments of the present invention do not limit this; that is, the embodiments of the present invention do not limit the specific structure of the accurate judgment agent. It should be noted that the implementation method of calling the first agent and inferring the task execution input data of each task according to the adjusted agent skill can be the same as the implementation method of calling the first agent and inferring the task execution input data of each task according to the initial agent skill, and the embodiments of the present invention will not repeat it here.

[0045] Optionally, in other embodiments, it can also be determined whether the inference output of any task conflicts with the inference output of other tasks. If there is a conflict, it can be determined that the inference output of any task is incorrect; if there is no conflict, it can be determined that the inference output of any task is accurate, and so on. The embodiments of the present invention do not limit this.

[0046] It should be noted that any two intelligent agents in the embodiments of the present invention may be the same or different, and the embodiments of the present invention do not limit this.

[0047] This invention provides an embodiment that can acquire initial agent skills corresponding to a target task type, as well as a training data set. The training data set includes training data for each task among multiple tasks under the target task type, with each training data set including the task execution input data for one task. A first agent is invoked to infer the task execution input data for each task according to the initial agent skills, obtaining the execution trajectory and inference output for each task. Then, the initial agent skills can be decomposed to obtain multiple sub-skills. Based on the execution trajectory and inference output of each task, the utility value of each sub-skill is determined. Further, based on the utility values ​​of each sub-skill, a sub-skill to be evolved can be determined from the multiple sub-skills. A second agent is invoked to adjust the sub-skill to be evolved, obtaining an adjusted agent skill. Correspondingly, the target agent skill for the target task type can be determined based on the adjusted agent skill. As can be seen, the embodiments of the present invention can adjust the sub-skills to be evolved in the initial intelligent agent skills through the intelligent agent, obtain the adjusted intelligent agent skills, and then determine the target intelligent agent skills with higher accuracy, thereby realizing the automated evolution of intelligent agent skills and effectively improving the accuracy of target intelligent agent skills. Furthermore, the embodiments of the present invention can make the evolution direction of intelligent agent skills clear through the sub-skills to be evolved, no longer blindly modifying them, thus changing from "global random modification" to "precise positioning of the sub-skills to be evolved", making the evolution process highly controllable, thereby effectively improving the evolution efficiency.

[0048] Based on the above description, this embodiment of the invention also proposes a more specific method for the evolution of intelligent agent skills. Accordingly, this method for the evolution of intelligent agent skills can be executed by the aforementioned electronic device (terminal or server); or, it can be executed jointly by the terminal and the server. For ease of explanation, the following description will use the execution of this method by an electronic device as an example; please refer to [link to previous text]. Figure 3 The method for evolving the skills of an intelligent agent may include the following steps S301-S307: S301, obtain the initial agent skills corresponding to the target task type, and obtain the training data set; the training data set includes the training data of each task in multiple tasks under the target task type, and each training data includes the task execution input data of one task.

[0049] S302, invoke the first intelligent agent, and according to the initial intelligent agent skills, reason about the task execution input data of each task to obtain the execution trajectory and reasoning output of each task.

[0050] S303, decompose the initial agent skill into multiple sub-skills.

[0051] Based on this, embodiments of the present invention can achieve fine-grained slicing (i.e. decomposition) of skills to segment into the smallest non-overlapping independent segments according to delimiters or semantics.

[0052] S304, for any one of the multiple sub-skills, determine the contribution value of any one sub-skill to each task based on the execution trajectory and inference output of each task.

[0053] Optionally, the contribution value of a sub-skill to a task can also be referred to as the contribution of the corresponding sub-skill to the corresponding task, and so on. Accordingly, the contribution value of a sub-skill to a task can be used to indicate the degree of contribution of the corresponding sub-skill to the corresponding task.

[0054] Optionally, for any one of multiple tasks, the electronic device may invoke a third intelligent agent to determine the contribution value of any sub-skill to any task based on the task execution input data, execution trajectory, inference output, and any sub-skill of that task. For example, the task execution input data, execution trajectory, inference output, and any sub-skill of any task can be input into the third intelligent agent to output the contribution value of any sub-skill to any task; or, the task execution input data, execution trajectory, inference output, result label, and any sub-skill of any task can be input into the third intelligent agent to output the contribution value of any sub-skill to any task; or, the task execution input data, execution trajectory, inference output, result label, any sub-skill, and contribution value calculation prompt can be input into the third intelligent agent to output the contribution value of any sub-skill to any task, and so on; this embodiment of the invention does not limit this.

[0055] Alternatively, for any one of multiple tasks, the electronic device can also invoke a third intelligent agent to determine the contribution value of each sub-skill to any task based on the task execution input data, execution trajectory, inference output, and each sub-skill of that task. Correspondingly, after obtaining the contribution values ​​of each sub-skill to each task, the contribution value of any sub-skill to each task can be determined, and so on. This embodiment of the invention does not limit this approach. For example, the task execution input data, execution trajectory, inference output, and each sub-skill of any task can be input into the third intelligent agent to output the contribution value of each sub-skill to any task; or, the task execution input data, execution trajectory, inference output, result label, and each sub-skill of any task can be input into the third intelligent agent to output the contribution value of each sub-skill to any task; or, the task execution input data, execution trajectory, inference output, result label, each sub-skill, and contribution value calculation prompt words of any task can be input into the third intelligent agent to output the contribution value of each sub-skill to any task, and so on. This embodiment of the invention does not limit this approach.

[0056] Optionally, the third intelligent agent can be any intelligent agent, and this embodiment of the invention does not limit this; that is, this embodiment of the invention does not limit the specific structure of the third intelligent agent. Optionally, the contribution value calculation prompt can be set according to experience or actual needs, and this embodiment of the invention does not limit this; for example, the contribution value calculation prompt can be calculating the contribution gain of the sub-skill to the task, etc.

[0057] Optionally, the contribution value of a sub-skill to a task can be used to indicate the contribution of the corresponding sub-skill to the corresponding task, that is, to indicate the benefit that the corresponding sub-skill brings to the corresponding task, and so on. Accordingly, the higher the contribution value, the higher the benefit, that is, the better the effect produced, and so on.

[0058] Optionally, the electronic device can also tag each of the influencing sub-skills in the set of influencing sub-skills for any task. An influencing sub-skill for a task can refer to a sub-skill that has an impact on the corresponding task, such as a skill whose contribution value to the corresponding task is not zero, and so on.

[0059] Optionally, in other embodiments, the electronic device may first invoke an intelligence agent to determine the set of influencing sub-skills for any task, based on the task execution input data, execution trajectory, inference output, and influencing sub-skill judgment prompts for any task. For example, the task execution input data, execution trajectory, inference output, and influencing sub-skill judgment prompts for any task can be input into the intelligence agent to output the set of influencing sub-skills for any task, etc. Then, a third intelligence agent may be invoked to determine the contribution value of each influencing sub-skill for any task to any task, based on the task execution input data, execution trajectory, inference output, and influencing sub-skills for any task, etc. This embodiment of the invention does not limit this. Optionally, the intelligence agent to determine the influencing sub-skills can be any intelligence agent, and this embodiment of the invention does not limit this; that is, this embodiment of the invention does not limit the specific structure of the intelligence agent to determine the influencing sub-skills. Optionally, the influencing sub-skill judgment prompts can be set according to experience or actual needs, and this embodiment of the invention does not limit this; for example, the influencing sub-skill judgment prompts can be all sub-skills that affect the task determined from multiple sub-skills, etc. Optionally, when invoking a third intelligent agent to determine the contribution value of each influencing sub-skill of any task to any task based on the task execution input data, execution trajectory, inference output, and the set of influencing sub-skills of any task, the task execution input data, execution trajectory, inference output, result label, and the set of influencing sub-skills of any task can be input into the third intelligent agent so that the third intelligent agent can output the contribution value of each influencing sub-skill of any task to any task; or, the task execution input data, execution trajectory, success or failure flag (such as a success flag or failure flag, which can be used to indicate whether it is successful) and the set of influencing sub-skills of any task can be input into the third intelligent agent so that the third intelligent agent can output the contribution value of each influencing sub-skill of any task to any task, and so on; the embodiments of the present invention do not limit this. It should be understood that the contribution value of each sub-skill other than the set of influencing sub-skills of any task to any task can be 0. Optionally, the success flag (which can be used to indicate success) and the failure flag (which can be used to indicate failure) can both be set according to experience or actual needs, and the embodiments of the present invention do not limit this.

[0060] Based on this, embodiments of the present invention can achieve sub-skill attribution for any task through the execution trajectory of any task, that is, determine the sub-skills that affect any sub-task, thereby locating the sub-skills that affect the reasoning output of any task.

[0061] S305, determine the utility value of any sub-skill based on its contribution to each task.

[0062] Optionally, a contribution value can be any value within the range of contribution values, and this embodiment of the present invention does not limit this; optionally, the range of contribution values ​​can be set according to experience or actual needs, and this embodiment of the present invention does not limit this; for example, the range of contribution values ​​can be from -1 to 1, etc.

[0063] In one implementation, the electronic device may also determine the difficulty value of each task; and determine the utility value of any sub-skill based on the contribution value of any sub-skill to each task and the difficulty value of each task.

[0064] Optionally, the electronic device can invoke a difficulty determination agent to determine the difficulty value of each task based on the task execution input data. For example, the task execution input data and / or difficulty determination prompts for each task can be input into the difficulty determination agent, which then outputs the difficulty value of each task. Alternatively, the task execution input data and / or difficulty determination prompts for each task can be input sequentially into the difficulty determination agent, which then outputs the difficulty value of each task sequentially, and so on. This embodiment of the invention does not limit this approach. Alternatively, the difficulty value of each task can be set according to experience or actual needs, and this embodiment of the invention does not limit this approach. For example, the difficulty value of each task can all be 1, in which case the differences between different tasks themselves can be disregarded, and so on. Optionally, the difficulty determination agent can be any agent, and this embodiment of the invention does not limit this approach. That is, this embodiment of the invention does not limit the specific structure of the difficulty determination agent. Optionally, the difficulty determination prompts can be set according to experience or actual needs, and this embodiment of the invention does not limit this approach. For example, the difficulty determination prompts can be used to determine the difficulty value of each task, and so on.

[0065] Optionally, when determining the utility value of any sub-skill based on its contribution to each task and the difficulty value of each task, the product of the sub-skill's contribution to each task and the difficulty value of each task can be used as the sub-skill's contribution to the difficulty of that task. Therefore, the average of the sub-skill's contribution to the difficulty of each task can be used as the utility value of that sub-skill, thus determining the utility value of any sub-skill. For example, let any sub-skill be considered as sub-skill P. k Taking an example, electronic devices can use Formula 1.1 (also known as the skill segment utility attribution formula) to determine the utility value of any sub-skill: Formula 1.1 Among them, SUA(P k ) can represent the utility value (also known as the overall utility score) of any sub-skill, N can represent the total number of tasks in multiple tasks, and C nVal(P) represents the difficulty value of the nth task out of multiple tasks. k (n) can represent the contribution value of any sub-skill to the nth task (e.g., it can be output by the aforementioned third agent). In this embodiment of the invention, the difficulty value of a task can be used to represent the difficulty level of the corresponding task; wherein, when the difficulty value of the first task is greater than the difficulty value of the second task, the task difficulty of the first task is higher than the task difficulty of the second task, and the first task and the second task are any two tasks among the plurality of tasks.

[0066] In another implementation, the electronic device can also perform a weighted summation of the contribution values ​​of any sub-skill to each task to obtain the utility value of any sub-skill. Optionally, the weight of each task can be set according to experience or actual needs; or, the weight of a task can also be the difficulty value of the corresponding task, etc.; this embodiment of the invention does not limit this.

[0067] Based on this, the embodiments of the present invention enable the utility of each sub-skill to be calculated, the attribution process to be transparent, and facilitate auditing and debugging; in other words, the embodiments of the present invention can achieve quantifiability, explainability, and traceability.

[0068] Optionally, in other embodiments, the electronic device may also invoke a fourth intelligent agent to determine the utility value of each sub-skill based on the execution trajectory and inference output of each task, as well as each sub-skill. For example, the execution trajectory and inference output of each task, as well as each sub-skill, can be output to the fourth intelligent agent so that the utility value of each sub-skill can be directly output through the fourth intelligent agent, etc.; the present invention does not limit this. It should be noted that the present invention does not limit the specific structure of the fourth intelligent agent.

[0069] S306, based on the utility value of each sub-skill, determine the sub-skill to be evolved from multiple sub-skills; and call the second agent to adjust the sub-skill to be evolved, thus obtaining the adjusted agent skill.

[0070] Optionally, training data may further include a result label for a task; that is, training data for a task may also include the result label for the corresponding task. Based on this, when invoking the second agent to adjust the sub-skill to be evolved and obtain the adjusted agent skill, the electronic device may invoke the second agent to adjust the sub-skill to be evolved based on the execution trajectory, inference output, and result label of each task, thereby obtaining the adjusted agent skill. The adjusted agent skill may include the evolved sub-skill corresponding to the sub-skill to be evolved (also referred to as the adjusted sub-skill of the sub-skill to be evolved) and all sub-skills other than the sub-skill to be evolved among multiple sub-skills. Optionally, the second agent may be any agent, and this embodiment of the invention does not limit it; that is, this embodiment of the invention does not limit the specific structure of the second agent.

[0071] For example, an electronic device can input the execution trajectory, inference output, result label, and sub-skill prompts to be evolved for each task (such as prompts to adjust which sub-skills in the initial agent's skills) and the initial agent's skills into a second agent, so as to adjust the agent's skills through the output of the second agent; or, based on the inference output and result label of each task, it can determine whether the inference output of each task is accurate (i.e., whether the inference output is consistent with the corresponding result label), and take the execution trajectory of the task with accurate inference output as the successful trajectory, and the execution trajectory of the task with inaccurate inference output as the failed trajectory, and then input the successful trajectory, failed trajectory, initial agent skills, and sub-skill prompts to be evolved into the second agent, so as to adjust the agent's skills through the output of the second agent; or, it can input from multiple In each task, all tasks related to the sub-skill to be evolved that influence the sub-skill set (i.e., a task related to a sub-skill to be evolved can be a task whose influence the sub-skill set includes the sub-skill to be evolved) are identified. The execution trajectory, inference output, result label, initial agent skill, and sub-skill prompt for each identified task related to the sub-skill to be evolved are input into a second agent to adjust the agent skill. Alternatively, the execution trajectory, inference output, result label, and sub-skill to be evolved for each task can be input into the second agent to output the corresponding evolved sub-skill, and this evolved sub-skill can be used to replace the sub-skill to be evolved in the initial agent skill to obtain the adjusted agent skill, etc. This embodiment of the invention does not limit this. Optionally, the electronic device can also input the skill identifier of the sub-skill to be evolved (such as the position of the sub-skill to be evolved in the initial agent skill, etc.) into the second agent to adjust the agent skill, etc. Optionally, the skill identifier can be set according to experience or actual needs, and this embodiment of the invention does not limit this. Optionally, the prompt words for the sub-skill to be evolved can be set according to experience or actual needs, and this embodiment of the invention does not limit this; optionally, the prompt words for the sub-skill to be evolved may include the skill identifier of the sub-skill to be evolved, etc. Based on this, this embodiment of the invention can constrain the agent to only modify the sub-skill to be evolved, keeping other skill fragments unchanged, in order to generate a new version of the skill (i.e., adjust the agent's skills, which can also be called skill adjustment, etc.).

[0072] Optionally, in other embodiments, the accuracy can be determined solely based on the inference output, such as determining whether the various inference outputs conflict. In this case, a second agent can be invoked to adjust the sub-skill to be evolved based on the execution trajectory and inference output of each task, thereby obtaining the adjusted agent skill. For example, the execution trajectory, inference output, initial agent skill, and prompts for the sub-skill to be evolved can be input into the second agent to adjust the agent skill through the output of the second agent, and so on. This invention does not limit this.

[0073] Based on this, embodiments of the present invention can target and optimize sub-skills to be evolved, such as targeting and optimizing the sub-skill with the weakest utility value, thereby significantly improving the iteration speed. For example, for the same improvement in accuracy, it can save 60%-90% of the number of iterations.

[0074] S307, Based on adjusting the agent's skills, determine the target agent's skills for the target task type.

[0075] To further illustrate the agent skill evolution method proposed in this embodiment of the invention, this embodiment uses a Chinese-to-English translation task as an example for illustrative explanation. The task to be solved (i.e., task information) can be the translation of short Chinese sentences into English, or the agent skill can be the translation of short Chinese sentences into English. For example, assuming the training dataset includes training data for each of the three tasks, the task execution input data for task n1 (which can also be called inputting Chinese) can be "The weather is nice today." (the corresponding result label can be "The weather is nice today."), the task execution input data for task n2 can be "I like reading books." (the corresponding result label can be "I like reading books."), and the task execution input data for task n3 can be "Let's go to the park tomorrow." (the corresponding result label can be "Let's go to the park tomorrow."). Furthermore, assuming the initial agent skill is "receiving Chinese sentences; translating Chinese words word by word; piecing together English words according to Chinese word order; directly outputting the piecing result," in this case, a result label can also be called a standard reference output or a correct translation.

[0076] Then, the first intelligent agent can be invoked to reason about the task execution input data of each task according to the initial intelligent agent skills, so as to obtain the execution trajectory and reasoning output of each task. For example, the execution trajectory of task n1 could be: receive a sentence - translate word for word: today, weather, very, good - concatenate according to Chinese word order, the inference output (i.e., the actual output) of task n1 could be "Today weather very good.", compared with the reference output (i.e. the result label): sentence structure error - result: failure (i.e., inaccurate); the execution trajectory of task n2 could be: receive a sentence - translate word for word: I, like, read book - concatenate according to Chinese word order, the inference output of task n2 could be "I likeread book.", compared with the reference output: grammar error - result: failure; the execution trajectory of task n3 could be: receive a sentence - translate word for word: tomorrow, together, go, park - concatenate according to Chinese word order, the inference output of task n3 could be "Tomorrow together go park.", compared with the reference output: grammar error - result: failure; based on this, all 3 tasks failed, and the current accuracy rate is 0%.

[0077] Furthermore, the initial agent skill can be decomposed to obtain multiple sub-skills. For example, assuming that at least one separator includes a semicolon, the skill can be decomposed according to the semicolon, into non-overlapping sub-skills that can be modified independently. At this time, the multiple sub-skills can include sub-skill P0, sub-skill P1, sub-skill P2, and sub-skill P3. Sub-skill P0 can be receiving Chinese sentences, sub-skill P1 can be translating Chinese words word by word, sub-skill P2 can be concatenating English words according to Chinese word order, and sub-skill P3 can be directly outputting the concatenated result. Correspondingly, each task can be attributed and scored separately to obtain the contribution value (also called contribution score) of each sub-skill for each task. Assuming the contribution value range is [-1, 1], a positive contribution (i.e., assisting in correctness) can have a contribution value of (0, 1), a negative contribution (i.e., causing an error) can have a contribution value of [-1, 0), and no impact can have a contribution value of 0. For example, when attributing and scoring task n1, the root cause of the error (i.e., the corresponding set of influencing sub-skills) can include word-for-word translation, Chinese word order concatenation, and sub-skill P0's contribution value for task n1 can be 0, sub-skill P1's contribution value for task n1 can be -0.8, sub-skill P2's contribution value for task n1 can be -0.9, and sub-skill P3's contribution value for task n1 can be -0.9. The contribution value of sub-skill P0 to task n1 can be 0. When attributing and scoring task n2, the root causes of errors may include word-for-word translation and splicing of Chinese word order. The contribution value of sub-skill P0 to task n2 can be 0, the contribution value of sub-skill P1 to task n2 can be -0.7, the contribution value of sub-skill P2 to task n2 can be -0.8, and the contribution value of sub-skill P3 to task n2 can be 0. When attributing and scoring task n3, the root causes of errors may include word-for-word translation and splicing of Chinese word order. The contribution value of sub-skill P0 to task n3 can be 0, the contribution value of sub-skill P1 to task n3 can be -0.75, the contribution value of sub-skill P2 to task n3 can be -0.85, and the contribution value of sub-skill P3 to task n3 can be 0.

[0078] Correspondingly, the utility value of any sub-skill can be determined based on its contribution to each task. For example, using Formula 1.1, and assuming the difficulty value of each task is 1, the utility values ​​of each sub-skill can be 0, -0.75, -0.85, and 0 respectively. In this case, taking the sub-skill to be evolved as the weakest utility segment (i.e., the sub-skill with the smallest utility value) as an example, the sub-skill to be evolved can be sub-skill P2. In this case, only sub-skill P2 can be adjusted, while sub-skills P0, P1, and P3 remain unchanged, thus obtaining the adjusted agent skill. For example, assuming the evolved sub-skill (i.e., the evolved sub-skill P2) is to recombine words according to English grammar and adjust sentence structure, the adjusted agent skill can be "receive Chinese sentences; translate Chinese words word for word; recombine words according to English grammar and adjust sentence structure; directly output the concatenated result."

[0079] Furthermore, the first agent can be invoked to infer the task execution input data of each task according to the adjusted agent's skills, thereby obtaining the execution trajectory and inference output of each task (i.e., the execution trajectory and inference output under the current skill, or the execution trajectory and inference output under the adjusted agent's skills, etc.). Based on the current inference output of each task (i.e., the inference output under the current skill), the accuracy is calculated to determine whether the evolution stopping condition is met. If the evolution stopping condition is met, the current adjusted agent skill is output; if the evolution stopping condition is not met, the iteration continues until the evolution stopping condition is met, and so on.

[0080] As can be seen, the embodiments of the present invention can realize a complete skill evolution system of "skill decomposition-attribution-quantification of utility value-locating sub-skills to be evolved-targeted optimization of sub-skills to be evolved-evaluation of whether the evolution stopping condition is met-convergence (i.e., the evolution stopping condition is met)".

[0081] This invention, after acquiring the initial agent skill corresponding to the target task type and the training data set, calls a first agent to infer the task execution input data of each task according to the initial agent skill, obtaining the execution trajectory and inference output of each task; and decomposes the initial agent skill into multiple sub-skills. Then, for any sub-skill among the multiple sub-skills, the contribution value of any sub-skill to each task can be determined based on the execution trajectory and inference output of each task; and the utility value of any sub-skill can be determined based on the contribution value of any sub-skill to each task. Accordingly, based on the utility values ​​of each sub-skill, a sub-skill to be evolved can be determined from the multiple sub-skills; and a second agent is called to adjust the sub-skill to be evolved, obtaining the adjusted agent skill. Furthermore, based on the adjusted agent skill, the target agent skill for the target task type can be determined. It is evident that this invention can modify only the sub-skill to be evolved without affecting other effective content (i.e., without destroying the original knowledge), thereby avoiding performance fluctuations and degradation, making skill evolution more stable, and thus effectively improving the accuracy of the target agent skill.

[0082] Based on the description of the relevant embodiments of the intelligent agent skill evolution method above, this invention also proposes an intelligent agent skill evolution device, which can be a computer program (including program code) running in an electronic device; such as Figure 4 As shown, the intelligent agent skill evolution device may include an acquisition unit 401 and a processing unit 402. The intelligent agent skill evolution device can perform... Figure 1 or Figure 3 The illustrated method for evolving agent skills, i.e., the agent skill evolution device, can operate the aforementioned unit: The acquisition unit 401 is used to acquire the initial agent skills corresponding to the target task type and to acquire the training data set; the training data set includes the training data of each task in multiple tasks under the target task type, and one training data includes the task execution input data of one task; Processing unit 402 is used to invoke the first intelligent agent to reason about the task execution input data of each task according to the initial intelligent agent skills, so as to obtain the execution trajectory and reasoning output of each task; The processing unit 402 is further configured to perform skill decomposition on the initial agent skill to obtain multiple sub-skills of the initial agent skill; and to determine the utility value of each sub-skill in the multiple sub-skills based on the execution trajectory and inference output of each task. The processing unit 402 is further configured to determine, based on the utility value of each sub-skill, a sub-skill to be evolved from the plurality of sub-skills, the sub-skill to be evolved including the sub-skill with the smallest utility value among the plurality of sub-skills; and to invoke a second agent to adjust the sub-skill to be evolved, thereby obtaining an adjusting agent skill; The processing unit 402 is further configured to determine the target agent skill of the target task type based on the adjusted agent skill.

[0083] In one embodiment, when the processing unit 402 performs skill decomposition on the initial agent skill to obtain multiple sub-skills of the initial agent skill, it may specifically be used to: Determine the skill decomposition strategy corresponding to the target task type, wherein the skill decomposition strategy is a separator segmentation strategy or a semantic segmentation strategy; According to the skill decomposition strategy, the initial agent skill is decomposed to obtain multiple sub-skills of the initial agent skill.

[0084] In another implementation, when determining the utility value of each sub-skill among the plurality of sub-skills based on the execution trajectory and inference output of each task, the processing unit 402 may specifically be used to: For any one of the multiple sub-skills, based on the execution trajectory and inference output of each task, determine the contribution value of any one sub-skill to each task; The utility value of any sub-skill is determined based on its contribution to each task.

[0085] In another implementation, when determining the contribution value of any sub-skill to each task based on the execution trajectory and inference output of each task, the processing unit 402 may specifically be used to: For any of the multiple tasks, a third intelligent agent is invoked to determine the contribution value of any sub-skill to any task based on the task execution input data, execution trajectory, inference output, and any sub-skill of the task.

[0086] In another embodiment, the processing unit 402 may also be used for: The difficulty value of each task is determined, and the difficulty value of a task is used to indicate the level of difficulty of the corresponding task; wherein, when the difficulty value of the first task is greater than the difficulty value of the second task, the task difficulty of the first task is higher than the task difficulty of the second task, and the first task and the second task are any two tasks among the plurality of tasks. When determining the utility value of any sub-skill based on its contribution to each task, the processing unit 402 may specifically be used to: The utility value of any sub-skill is determined based on its contribution to each task and the difficulty value of each task.

[0087] In another implementation, training data also includes a result label for a task; when the processing unit 402 invokes the second agent to adjust the sub-skill to be evolved, and obtains the adjusted agent skill, it can specifically be used for: The second intelligent agent is invoked to adjust the sub-skill to be evolved based on the execution trajectory, inference output and result label of each task, thereby obtaining the adjusted intelligent agent skill; The adjusted agent skills include the evolved sub-skill corresponding to the sub-skill to be evolved and all sub-skills other than the sub-skill to be evolved among the multiple sub-skills.

[0088] In another embodiment, when determining the target agent skill of the target task type based on the adjusted agent skill, the processing unit 402 may specifically be used to: The adjusted agent skill is used as the initial agent skill, and the first agent is called iteratively executed. According to the initial agent skill, the task execution input data of each task is reasoned to achieve continuous evolution of the agent skill until the evolution stop condition is met. The evolution stop condition is used to determine whether to terminate the evolution iteration. The adjusted agent skill that meets the evolution stopping condition will be used as the target agent skill for the target task type.

[0089] According to one embodiment of the present invention, Figure 4 Each unit in the illustrated agent skill evolution device can be individually or entirely merged into one or more other units, or one or more of the units can be further divided into multiple functionally smaller units. This achieves the same operation without affecting the technical effects of the embodiments of the present invention. The above units are based on logical function division. In practical applications, the function of one unit can be implemented by multiple units, or the function of multiple units can be implemented by one unit. In other embodiments of the present invention, any agent skill evolution device may also include other units. In practical applications, these functions can also be implemented with the assistance of other units, and can be implemented collaboratively by multiple units.

[0090] According to another embodiment of the present invention, it is possible to perform operations such as those described above by running on a general-purpose electronic device, such as a computer, which includes processing elements and storage elements such as a central processing unit (CPU), random access memory (RAM), and read-only memory (ROM). Figure 1 or Figure 3The computer program (including program code) involved in each step of the corresponding method shown, to construct such... Figure 4 The diagram illustrates an agent skill evolution device and a method for implementing agent skill evolution according to embodiments of the present invention. The computer program can be stored on, for example, a computer storage medium, loaded onto the aforementioned electronic device via the computer storage medium, and run therein.

[0091] Based on the description of the method and apparatus embodiments above, an exemplary embodiment of the present invention also provides an electronic device, including: at least one processor; and a memory communicatively connected to the at least one processor. The memory stores a computer program executable by the at least one processor, which, when executed by the at least one processor, causes the electronic device to perform the method according to an embodiment of the present invention.

[0092] An exemplary embodiment of the present invention also provides a non-transitory computer-readable storage medium storing a computer program, wherein the computer program, when executed by a computer's processor, is used to cause the computer to perform a method according to an embodiment of the present invention.

[0093] An exemplary embodiment of the present invention also provides a computer program product, including a computer program, wherein, when executed by a computer's processor, the computer program is used to cause the computer to perform a method according to an embodiment of the present invention.

[0094] refer to Figure 5 The present invention will now be described in the form of a structural block diagram of an electronic device 500 that can serve as a server or client of the present invention, which is an example of a hardware device that can be applied to various aspects of the present invention. The electronic device is intended to represent various forms of digital electronic computer devices, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the invention described and / or claimed herein.

[0095] like Figure 5As shown, the electronic device 500 includes a computing unit 501, which can perform various appropriate actions and processes according to a computer program stored in a read-only memory (ROM) 502 or a computer program loaded from a storage unit 508 into a random access memory (RAM) 503. The RAM 503 may also store various programs and data required for the operation of the electronic device 500. The computing unit 501, ROM 502, and RAM 503 are interconnected via a bus 504. An input / output (I / O) interface 505 is also connected to the bus 504.

[0096] Multiple components in electronic device 500 are connected to I / O interface 505, including: input unit 506, output unit 507, storage unit 508, and communication unit 509. Input unit 506 can be any type of device capable of inputting information to electronic device 500. Input unit 506 can receive input digital or character information and generate key signal inputs related to user settings and / or function control of electronic device. Output unit 507 can be any type of device capable of presenting information and may include, but is not limited to, a display, speaker, video / audio output terminal, vibrator, and / or printer. Storage unit 508 may include, but is not limited to, disk and optical disk. Communication unit 509 allows electronic device 500 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks, and may include, but is not limited to, modems, network cards, infrared communication devices, wireless communication transceivers, and / or chipsets, such as Bluetooth™ devices, WiFi devices, WiMax devices, cellular communication devices, and / or the like.

[0097] The computing unit 501 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 501 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 501 performs the various methods and processes described above. For example, in some embodiments, the agent skill evolution method can be implemented as a computer software program tangibly contained in a machine-readable medium, such as storage unit 508. In some embodiments, part or all of the computer program can be loaded and / or installed on the electronic device 500 via ROM 502 and / or communication unit 509. In some embodiments, the computing unit 501 can be configured to perform the agent skill evolution method by any other suitable means (e.g., by means of firmware).

[0098] The program code used to implement the methods of the present invention can be written in any combination of one or more programming languages. This program code can be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing device, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code can be executed entirely on the machine, partially on the machine, as a standalone software package partially on the machine and partially on a remote machine, or entirely on a remote machine or server.

[0099] In the context of this invention, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. Machine-readable media can include, but are not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0100] As used herein, the terms "machine-readable medium" and "computer-readable medium" refer to any computer program product, device, and / or apparatus (e.g., disk, optical disk, memory, programmable logic device (PLD)) for providing machine instructions and / or data to a programmable processor, including machine-readable media that receive machine instructions as machine-readable signals. The term "machine-readable signal" refers to any signal for providing machine instructions and / or data to a programmable processor.

[0101] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0102] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as a data server), or computing systems that include middleware components (e.g., an application server), or computing systems that include frontend components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., a communication network). Examples of communication networks include local area networks (LANs), wide area networks (WANs), and the Internet.

[0103] Computer systems can include clients and servers. Clients and servers are generally located far apart and typically interact through communication networks. Client-server relationships are created by computer programs running on the respective computers and having a client-server relationship with each other.

[0104] Furthermore, it should be understood that the above-disclosed embodiments are merely preferred embodiments of the present invention and should not be construed as limiting the scope of the present invention. Therefore, any equivalent variations made in accordance with the claims of the present invention are still within the scope of the present invention.

Claims

1. A method for the evolution of intelligent agent skills, characterized in that, include: Obtain the initial agent skills corresponding to the target task type, and obtain the training data set; The training dataset includes training data for each task among multiple tasks under the target task type, and each training dataset includes the task execution input data for one task. The first intelligent agent is invoked to reason about the task execution input data of each task according to the initial intelligent agent skills, so as to obtain the execution trajectory and reasoning output of each task; The initial agent skill is decomposed to obtain multiple sub-skills of the initial agent skill; Based on the execution trajectory and inference output of each task, the utility value of each sub-skill among the multiple sub-skills is determined respectively; Based on the utility value of each sub-skill, a sub-skill to be evolved is determined from the plurality of sub-skills, and the sub-skill to be evolved includes the sub-skill with the smallest utility value among the plurality of sub-skills; And then invoke a second intelligent agent to adjust the sub-skill to be evolved, thus obtaining the adjusted intelligent agent skill; Based on the adjusted agent skills, the target agent skills for the target task type are determined.

2. The method according to claim 1, characterized in that, The process of decomposing the initial agent skill to obtain multiple sub-skills of the initial agent skill includes: Determine the skill decomposition strategy corresponding to the target task type, wherein the skill decomposition strategy is a separator segmentation strategy or a semantic segmentation strategy; According to the skill decomposition strategy, the initial agent skill is decomposed to obtain multiple sub-skills of the initial agent skill.

3. The method according to claim 1 or 2, characterized in that, The determination of the utility value of each sub-skill among the plurality of sub-skills based on the execution trajectory and inference output of each task includes: For any one of the multiple sub-skills, based on the execution trajectory and inference output of each task, determine the contribution value of any one sub-skill to each task; The utility value of any sub-skill is determined based on its contribution to each task.

4. The method according to claim 3, characterized in that, The determination of the contribution value of any sub-skill to each task based on the execution trajectory and inference output of each task includes: For any of the multiple tasks, a third intelligent agent is invoked to determine the contribution value of any sub-skill to any task based on the task execution input data, execution trajectory, inference output, and any sub-skill of the task.

5. The method according to claim 3, characterized in that, The method further includes: The difficulty value of each task is determined, and the difficulty value of a task is used to indicate the level of difficulty of the corresponding task; wherein, when the difficulty value of the first task is greater than the difficulty value of the second task, the task difficulty of the first task is higher than the task difficulty of the second task, and the first task and the second task are any two tasks among the plurality of tasks. The determination of the utility value of any sub-skill based on its contribution to each task includes: The utility value of any sub-skill is determined based on its contribution to each task and the difficulty value of each task.

6. The method according to claim 1 or 2, characterized in that, Training data also includes the result labels for a task; The process of invoking the second intelligent agent to adjust the sub-skill to be evolved, resulting in the adjusted intelligent agent skill, includes: The second intelligent agent is invoked to adjust the sub-skill to be evolved based on the execution trajectory, inference output and result label of each task, thereby obtaining the adjusted intelligent agent skill; The adjusted agent skills include the evolved sub-skill corresponding to the sub-skill to be evolved and all sub-skills other than the sub-skill to be evolved among the multiple sub-skills.

7. The method according to claim 1 or 2, characterized in that, The target agent skill for determining the target task type based on the adjusted agent skill includes: The adjusted agent skill is used as the initial agent skill, and the first agent is called iteratively executed. According to the initial agent skill, the task execution input data of each task is reasoned to achieve continuous evolution of the agent skill until the evolution stop condition is met. The evolution stop condition is used to determine whether to terminate the evolution iteration. The adjusted agent skill that meets the evolution stopping condition will be used as the target agent skill for the target task type.

8. A device for the skill evolution of an intelligent agent, characterized in that, The device includes: The acquisition unit is used to acquire the initial agent skills corresponding to the target task type and to acquire the training data set; the training data set includes the training data of each task in multiple tasks under the target task type, and each training data includes the task execution input data of one task; The processing unit is used to invoke the first intelligent agent to reason about the task execution input data of each task according to the initial intelligent agent skills, so as to obtain the execution trajectory and reasoning output of each task; The processing unit is further configured to decompose the initial agent skill into multiple sub-skills; and based on the execution trajectory and inference output of each task, determine the utility value of each sub-skill among the multiple sub-skills. The processing unit is further configured to determine, based on the utility value of each sub-skill, a sub-skill to be evolved from the plurality of sub-skills, the sub-skill to be evolved including the sub-skill with the smallest utility value among the plurality of sub-skills; and to invoke a second agent to adjust the sub-skill to be evolved, thereby obtaining an adjusting agent skill; The processing unit is further configured to determine the target agent skill of the target task type based on the adjusted agent skill.

9. An electronic device, characterized in that, include: processor; as well as Stored program memory, The program includes instructions that, when executed by the processor, cause the processor to perform the method according to any one of claims 1-7.

10. A non-transitory computer-readable storage medium storing computer instructions, characterized in that, The computer instructions are used to cause the computer to perform the method according to any one of claims 1-7.

Images