Intelligent agent fault attribution method and device based on a computational impedance model, and a medium

Abstract

The application discloses an agent fault attribution method and device based on a computing impedance model and a medium, and relates to the technical field of computers. The method comprises the following steps: obtaining traceable logs generated by an agent performing a target task, and constructing a circuit topology based on the traceable logs; determining local computing impedance corresponding to each node in the circuit topology by using a computing impedance model according to the decision confidence, action complexity and knowledge defect degree corresponding to each node; screening out a target node according to the local computing impedance, determining the cause of the agent fault according to the impedance composition of the target node, and generating corresponding optimization instructions for updating the agent. By mapping the task execution process of the agent to the circuit topology and using the impedance model, the decision confidence, action complexity and knowledge defect degree involved in the agent are quantified as impedance in the circuit, thereby improving the accuracy of agent fault analysis and the efficiency and reliability of agent optimization.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer, in particular to an agent fault attribution method and device based on a computational impedance model and a medium. BACKGROUND

[0002] Existing agent systems based on large language models (LLM), such as agents deployed in autonomous software development, DevOps (Development and Operations) chain diagnosis, or complex decision-making environments, face fundamental challenges in achieving continuous self-improvement and adapting to dynamic environments.

[0003] In related technologies, the fault detection of the agent has defects, which will limit the ability of the agent to evolve autonomously. The self-reflection mechanism used in related technologies usually relies on qualitative inspection of the logic or semantics of the chain of thought (CoT), which is essentially a correlation diagnosis. Only judging errors based on semantic coherence lacks the causal constraints of the physical world and cannot effectively separate surface errors from cross-level causal roots (such as being able to detect that the code written by the agent cannot run, but unable to determine the root cause of the error, i.e., knowledge failure, planning reasoning failure, or tool failure), and is easily affected by the output bias of the LLM itself and illusory causal links, resulting in a diagnostic result that is far less accurate and reliable than actual application requirements; for example, in a multi-agent system (MAS), identifying the steps that led to task failure and decisive errors still cannot be detected. SUMMARY

[0004] Therefore, the purpose of the present application is to provide an agent fault attribution method and device based on a computational impedance model, which can improve the accuracy of agent fault analysis. The specific solutions are as follows:

[0005] In a first aspect, the present application discloses an agent fault attribution method based on a computational impedance model, comprising:

[0006] Obtaining a traceable log generated by an agent executing a target task, and constructing a circuit topology based on the traceable log; one node in the circuit topology corresponds to one decision in a decision sequence of the target task, and the connection relationship between nodes in the circuit topology represents the logical relationship between different decisions;

[0007] According to the decision confidence, action complexity, and knowledge defect degree of each node in the circuit topology, the local computational impedance of each node is determined using a computational impedance model.

[0008] According to the local computing impedance, a target node is screened out, a reason for the agent failure is determined according to the impedance composition of the target node, and an optimization instruction for updating the agent is generated.

[0009] Optionally, the local computing impedance corresponding to each node in the circuit topology is determined by using a computing impedance model according to the decision confidence, action complexity and knowledge defect degree corresponding to each node.

[0010] The local computing impedance corresponding to the current node is obtained by weighted calculation based on the decision confidence corresponding to the current node, the entropy of the action sequence performed by the current node, and the difference degree between the new knowledge retrieved by the current node and the old knowledge in the agent, and the weights corresponding to each item.

[0011] The decision confidence and the local computing impedance are negatively correlated, the entropy and the local computing impedance are positively correlated, and the difference degree and the local computing impedance are positively correlated.

[0012] Optionally, after determining the local computing impedance corresponding to each node, the method further comprises:

[0013] The reference total computing impedance corresponding to the circuit topology is calculated.

[0014] The actual total computing impedance corresponding to the circuit topology is calculated by circuit analysis according to the local computing impedance of each node and the circuit topology.

[0015] The weights of the decision confidence, the entropy and the difference degree are updated according to the difference degree between the reference total computing impedance and the actual total computing impedance.

[0016] Optionally, the reference total computing impedance corresponding to the circuit topology is calculated by:

[0017] A target parameter serving as a task success score is determined according to the task type of the target task.

[0018] The task failure severity is determined based on the task success score, and the computing throughput is determined according to the effective information in the agent task execution process.

[0019] The total computing impedance corresponding to the task is calculated by taking the task failure severity as the potential difference and taking the computing throughput as the effective current.

[0020] Optionally, before determining the computing throughput according to the effective information in the agent task execution process, the method further comprises:

[0021] The content generated by the intelligent agent is subjected to syntax analysis, and the content meeting syntactic validity is taken as the valid information;

[0022] The knowledge searched by the intelligent agent is subjected to relevance threshold checking, and the knowledge passing the checking is taken as the valid information;

[0023] The number of messages in the conflict-free delivery is taken as the valid information.

[0024] Optionally, the target node is screened out according to the local computing impedance, and the screening comprises:

[0025] The node corresponding to the maximum local computing impedance is taken as the target node, or the N nodes with the highest local computing impedance values are taken as the target nodes, or the node with the local computing impedance greater than a preset local computing impedance threshold is taken as the target node.

[0026] Optionally, the traceable log comprises cognitive data and execution data, the cognitive data comprises a thinking chain and a tool calling sequence, and the execution data comprises a stack trace and an error code related to running.

[0027] Optionally, the fault cause of the intelligent agent is determined according to the impedance constitution of the target node, and an optimization instruction for updating the intelligent agent is generated, and the determining comprises:

[0028] If the contribution degree of the knowledge defect degree of the target node to the local computing impedance is greater than a contribution degree threshold, a first optimization instruction is generated.

[0029] According to the first optimization instruction, a new weight of a target old knowledge point is determined based on a weight of the target old knowledge point and a forgetting rate parameter, or the target old knowledge point is isolated, and the target old knowledge point is a knowledge point related to the knowledge defect degree in the knowledge graph of the intelligent agent.

[0030] If the contribution degree of the action complexity of the target node to the local computing impedance is greater than a contribution degree threshold, a second optimization instruction is generated.

[0031] According to the second optimization instruction, a sequence of actions performed by the target node is disassembled into a sequence of atomic operations, a parameter constraint mode of tool calling is started, and an application program interface is called in a sequence from high to low priority, and the higher the priority, the simpler the function of the called application program interface, and the higher the priority, the higher the certainty.

[0032] If the contribution degree of the decision confidence of the target node to the local computing impedance is greater than a contribution degree threshold, a third optimization instruction is generated.

[0033] According to the third optimization instruction, the thought chain planning framework is updated, and / or a logical check rule for a decision in the decision sequence is added, and / or an anti-fact reasoning verification mechanism is increased.

[0034] In a second aspect, the application discloses an agent fault attribution device based on a computational impedance model, which comprises:

[0035] A circuit topology construction module is configured to acquire traceable logs generated by an agent in executing a target task, and construct a circuit topology based on the traceable logs; one node in the circuit topology corresponds to one decision in a decision sequence of the target task, and a connection relationship between nodes in the circuit topology represents a logical relationship between different decisions.

[0036] A local computational impedance calculation module is configured to determine a local computational impedance corresponding to each node according to a decision confidence, an action complexity and a knowledge defect degree corresponding to each node in the circuit topology by using a computational impedance model.

[0037] A fault cause determination module is configured to screen out a target node according to the local computational impedance, determine an agent fault cause according to an impedance composition of the target node, and generate a corresponding optimization instruction for updating the agent.

[0038] In a third aspect, the application discloses a computer readable storage medium for storing a computer program; wherein the computer program is executed by a processor to implement the agent fault attribution method based on the computational impedance model.

[0039] In the application, traceable logs generated by an agent in executing a target task are acquired, and a circuit topology is constructed based on the traceable logs; one node in the circuit topology corresponds to one decision in a decision sequence of the target task, and a connection relationship between nodes in the circuit topology represents a logical relationship between different decisions; a local computational impedance corresponding to each node is determined according to a decision confidence, an action complexity and a knowledge defect degree corresponding to each node in the circuit topology by using a computational impedance model; a target node is screened out according to the local computational impedance, an agent fault cause is determined according to an impedance composition of the target node, and a corresponding optimization instruction for updating the agent is generated. As can be seen from the above, by mapping the task execution process of the agent to the circuit topology and using the impedance model, the decision confidence, the action complexity and the knowledge defect degree involved in the decision of the agent are quantified as the impedance in the circuit, so that the agent fault cause is determined by analyzing the impedance, and the optimization instruction for updating the agent is generated according to the fault cause, thereby improving the accuracy of the agent fault analysis and the efficiency and reliability of the agent optimization. BRIEF DESCRIPTION OF DRAWINGS

[0040] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the related art, the drawings needed to be used in the embodiments or related art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of the provided drawings.

[0041] Figure 1 A flow chart of an agent fault attribution method based on a computational impedance model is provided for the present application.

[0042] Figure 2 A structural schematic diagram of an agent fault attribution device based on a computational impedance model is provided for the present application.

[0043] Figure 3 A structural diagram of an electronic device is provided for the present application. DETAILED DESCRIPTION

[0044] In order to make the purpose, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0045] In the related art, the self-reflection mechanism usually relies on qualitative inspection of the logic or semantics of the thinking chain, which is essentially a correlation diagnosis. Errors are only judged according to semantic coherence, which lacks causal constraints of the physical world, cannot effectively separate surface errors and cross-level causal roots, and is easily affected by output bias and illusory causal links of LLM itself, resulting in a much lower accuracy and reliability of the diagnosis result than the actual application requirement. To overcome the above technical problems, the present application proposes an agent fault attribution method based on a computational impedance model, which formalizes the agent execution process as a circuit and uses a computational impedance model to quantitatively and accurately attribute the root cause of complex and multi-granularity task failure, significantly improving the fault analysis accuracy of the agent system and the autonomous learning efficiency and reliability.

[0046] The embodiments of the present application disclose an agent fault attribution method based on a computational impedance model. The computational impedance model identifies a physical impedance model based on a circuit. Referring to Figure 1 The method can include the following steps:

[0047] Step S11: Obtain traceable logs generated by the agent in executing the target task, and construct a circuit topology based on the traceable logs; a node in the circuit topology corresponds to a decision in the decision sequence of the target task, and the connection relationship between nodes in the circuit topology represents the logical relationship between different decisions.

[0048] First, acquire traceable logs generated by the agent during the execution of a task. These traceable logs include cognitive data and execution data. The cognitive data includes thought chains and tool call sequences, while the execution data includes stack traces and runtime error codes. It is understood that the first step is to acquire traceable logs generated by the self-evolving agent during the execution of complex tasks. These logs contain high-level cognitive data, such as the agent's thought chains (CoT) and tool call sequences; they also include low-level execution data, such as stack traces (a call path report automatically generated by the system when an exception or error occurs during program execution, used to show the order of function calls when the error occurs), runtime error codes, etc.

[0049] Based on traceable logs, the task execution path of the intelligent agent is formalized into a circuit topology, which consists of a series of computing nodes ( The system is composed of nodes, each corresponding to a decision in the agent's decision sequence; such as a decision in knowledge retrieval, a decision in planning, a decision in code generation, or a decision in tool invocation. The connection relationships between nodes are defined according to the logical or temporal dependencies of the task, forming series, parallel, or hybrid circuit structures. For example, parallel decision steps correspond to parallel circuits, while time-dependent decision steps correspond to series circuits. This circuit topology achieves causal decoupling at two levels of granularity. On the one hand, it achieves decoupling at the macroscopic topological granularity level, i.e., by locating the temporal bottleneck in the task execution process through the series and parallel structure of circuit nodes; that is, identifying which step went wrong. On the other hand, it achieves decoupling at the microscopic cognitive granularity level, by subsequently calculating the local impedance of nodes to decouple the cognitive root cause of the failure of that decision step.

[0050] Step S12: Based on the decision confidence, action complexity, and knowledge deficiency degree of each node in the circuit topology, determine the local computational impedance of each node using the computational impedance model.

[0051] In this embodiment, for each node i in the circuit topology, the calculated impedance Zi of the node (i.e. the local calculated impedance mentioned above) is calculated. The local calculated impedance is a quantitative measure of the resistance or interference of information flow in the decision-making step corresponding to the node. Specifically, it can be analyzed based on three dimensions: the planning uncertainty, the action complexity, and the knowledge deficiency of the decision.

[0052] In some embodiments, the local computational impedance of each node is determined by a computational impedance model based on the decision confidence of the current node, the entropy of the action sequence performed by the current node, and the difference between the new knowledge retrieved by the current node and the old knowledge in the agent.

[0053] In a preferred embodiment, the formula of the computational impedance model is as follows:

[0054] ;

[0055] wherein, , , are trainable regularization coefficients used to dynamically adjust the weights of each impedance term, satisfying + + =1. represents the decision confidence of node i, represents the entropy of the action sequence of node i, represents the KL divergence between the new knowledge retrieved by node i and the old knowledge in the agent. Of course, in addition to using KL divergence to represent the difference, other methods can also be used.

[0056] represents the rule uncertainty impedance, which is used to quantitatively measure the risk of reasoning planning errors, is the internal confidence of the agent in its own decision when performing step i. The peak value of the Action Softmax probability of the large language model output can be extracted, or the large language model can be required to report its decision confidence in the form of a value within an interval through structured prompt engineering. The decision confidence is negatively correlated with the local computational impedance, so when the agent shows extremely low confidence in a key decision (i.e. tends to zero), the local computational impedance will increase exponentially, thus effectively indicating the uncertainty error in the planning or logic, which is an important step to identify the weak points of the agent's mind.

[0057] represents the action complexity impedance, which can specifically use the Shannon entropy measures the complexity and randomness of the generated action sequence. A high entropy value corresponds to a high inherent risk, as complex actions are more likely to introduce logical defects or execution failures. Through the action complexity impedance, the agent can be punished for generating unnecessary complexity or redundant operations, guiding the agent to adopt a more concise and robust planning path.

[0058] represents the knowledge timeliness impedance, which is used to quantitatively measure the degree of knowledge defects. Specifically, the Kullback-Leibler divergence (KL divergence) can be used to represent the difference between the new knowledge retrieved by the current node and the old knowledge in the agent. The KL divergence is an asymmetric measure of the difference between two probability distributions, which can quantify the difference between the new knowledge distribution retrieved at step i and the old knowledge distribution inherent to the agent . When the KL value is high, it indicates that there is a significant difference between the new and old knowledge, for example, the retrieved new API document conflicts with the stored old API document, the information flow is hindered, and the impedance will increase. By axiomatizing and quantitatively measuring knowledge defects, the long-standing confusion between knowledge defects and planning failures in traditional attribution methods is solved, significantly improving the targeting of fault attribution.

[0059] In some embodiments, after determining the local computing impedance corresponding to each node, the method further includes: calculating a reference total computing impedance corresponding to the circuit topology; calculating an actual total computing impedance corresponding to the circuit topology through circuit analysis according to the local computing impedance of each node and the circuit topology; and updating the weights of the decision confidence, the entropy, and the difference degree according to the difference between the reference total computing impedance and the actual total computing impedance. That is, the reference total computing impedance is an ideal value, and the values of the above , , are adjusted inversely according to the difference between the reference total computing impedance and the actual total computing impedance, improving the sensitivity of the agent to faults in order to improve the detection rate of faults. In addition, by performing computational analysis on the circuit, the actual total computing impedance is obtained according to the circuit connection rules (such as the sum of the impedances of series-connected nodes).

[0060] The computing the reference total computing impedance corresponding to the circuit topology comprises: determining a target parameter as a task success score according to a task type of the target task; determining a task failure severity based on the task success score, determining a computing throughput according to effective information in a task execution process of the agent; and taking the task failure severity as an electric potential difference and taking the computing throughput as an effective current to compute a total computing impedance corresponding to the task. In some embodiments, before the determining the computing throughput according to the effective information in the task execution process of the agent, the method further comprises: performing syntax analysis on content generated by the agent, and taking content meeting syntax validity as the effective information; performing relevance threshold verification on knowledge searched by the agent, and taking knowledge passing the verification as the effective information; and taking a number of conflict-free delivered messages as the effective information.

[0061] The final failure potential of the task The computing throughput in the execution process of the agent is associated, so as to determine the reference total computing impedance causing the failure. The failure phenomenon in the task execution process of the agent is mapped into a physical model of Ohm's law, so as to establish a computable causal constraint. The specific definition is as follows: the final failure potential of the task is formally analogous to an electric potential difference (voltage) in a circuit, and a calculation formula thereof is:

[0062] ;

[0063] wherein, a task success score (Task Success Score), is a normalized scalar (with a value range [0, 1]) for measuring a task target achievement degree. A specific calculation index thereof can be determined according to a task type; for example, for a code generation task, a pass rate (Pass Rate) or an execution success rate of a unit test case can be adopted; for a knowledge question and answer or reasoning task, an exact match degree (Exact Match) or an F1 score of a generated answer and a standard true value can be adopted; for an open planning task, a completion rate of a key sub-target or a human preference score (Human Preference Score) based on human feedback can be adopted. As a scalar, the macro signal driving the attribution analysis is represented, and the higher the value is, the more serious the failure is, that is, the greater the electric potential difference is.

[0064] The effective information amount successfully processed, generated or delivered by the agent in an execution path is defined as a computing throughput , and the computing throughput This refers to the throughput of effective information flow, which can be approximated by counting the number of lines of code successfully executed by the Agent, the number of effectively retrieved knowledge fragments, or the number of conflict-free messages transmitted. That is, its calculation not only counts the generated Tokens (the basic unit of text processing, i.e., the smallest unit when the AI ​​model processes text) or lines of code, but also incorporates analyses such as syntactic validity filters. For example, only generated code blocks that pass static syntax analysis, or retrieved knowledge fragments that pass a relevance threshold check, are included in the calculation. The effective current in the circuit is simulated by the effective information flow, eliminating interference from heat loss (i.e., invalid generation). The throughput is formally analogized to the current in the circuit, and the total calculated impedance is referenced.

[0065] ;

[0066] in, This indicates a direct proportional relationship. The reference total calculated impedance represents the agent's cumulative resistance or interference with the information flow throughout the task execution process. This constraint ensures that the attribution analysis is physically consistent; that is, the total impedance is the potential to lead to failure. The underlying cause of the increase is thus transformed from a vague concept of failure into a calculable impedance value, providing a basis for subsequent precise attribution.

[0067] Step S13: Filter out target nodes based on the local calculated impedance, determine the cause of agent failure based on the impedance composition of the target nodes, and generate corresponding optimization instructions for agent update.

[0068] The step of selecting target nodes based on the local calculated impedance includes: selecting the node corresponding to the maximum local calculated impedance as the target node; or selecting the top N nodes with the highest local calculated impedance values ​​as the target nodes; or selecting nodes with local calculated impedance values ​​greater than a preset local calculated impedance threshold as the target nodes. It can be understood that the node corresponding to the maximum local calculated impedance is considered the path with the highest impedance, and the top N nodes or nodes with impedance values ​​greater than the preset local calculated impedance threshold are paths with higher impedance. Subsequent targeted agent updates are performed based on the impedance composition of these paths, thereby achieving efficient autonomous evolution. Specifically, optimization instructions can be generated based on the impedance composition of each target node, thereby eliminating multiple high-impedance sources in parallel during a single evolution iteration, achieving a reduction in the actual total impedance of the system and a significant improvement in its capabilities.

[0069] The optimization method is determined by calculating the relative contribution of each impedance (planning uncertainty impedance, action complexity impedance, and knowledge deficiency impedance) in the local computational impedance. The relative contribution can be defined as the proportion of a certain impedance to the overall performance of the local computational impedance. Percentage of the total value.

[0070] In some embodiments, if the contribution of the knowledge defect level of the target node to the local computational impedance is greater than the contribution threshold, a first optimization instruction is generated; according to the first optimization instruction, a new weight of the target old knowledge point is determined based on the weight of the target old knowledge point and the forgetting rate parameter, or the target old knowledge point is isolated; the target old knowledge point is a knowledge point in the agent knowledge graph that is related to the knowledge defect level.

[0071] That is, if knowledge timeliness is a resistance Contribution of the item Exceeding the preset threshold (For example, 70%), then optimization instructions are generated to address knowledge deficiencies. The first optimization instruction includes: guiding the knowledge management module (e.g., a continuous learning module based on a graph database) to perform weight reduction or isolation (forgetting) operations on old knowledge nodes in the knowledge graph that are related to the KL conflict. The weight calculation formula is as follows:

[0072] ;

[0073] in, This represents the new weight of the old knowledge node. This represents the old weight of the old knowledge node; This is the forgetting rate parameter. Updating knowledge point weights ensures accurate forgetting, preventing outdated knowledge or incorrect experiences from interfering with subsequent decisions.

[0074] In some embodiments, if the contribution of the action complexity of the target node to the local computational impedance is greater than a contribution threshold, a second optimization instruction is generated; according to the second optimization instruction, the action sequence executed by the current node is decomposed into an atomic operation sequence, and / or, the parameter constraint mode of the tool call is enabled, and / or, the application programming interface is called in order of priority from high to low; wherein, the simpler the function of the application programming interface, the higher the priority, and the higher the determinism, the higher the priority.

[0075] That is, if the action complexity resistance If the relative contribution of an item is dominant, the generated second optimization instructions include: 1. Atomization Restructuring: guiding the planning module to decompose the current high-complexity action into a set of low-entropy atomic operation sequences; 2. Parameter Constraints: forcibly enabling parameter constraint mode during tool calls to limit the model from generating unnecessary redundant parameters; each tool has a strict mode definition when provided to the model, and even if the tool mode definition defines certain parameters as not necessary, the model may still use them when generating calls. Parameter constraints can reduce such parameters; 3. API Degradation and Replacement: guiding the Agent to prioritize searching for and calling alternative APIs with simpler functions and higher determinism to reduce the risk of randomness at the execution level.

[0076] In some embodiments, if the contribution degree of the decision confidence of the target node to the local computing impedance is greater than a contribution degree threshold, a third optimization instruction is generated; according to the third optimization instruction, the thinking chain planning framework is updated, and / or a logical verification rule for the decision in the decision sequence is added, and / or a counterfactual reasoning verification mechanism is added.

[0077] That is, if the contribution degree of the planning uncertainty impedance term dominates, an optimization execution for planning is generated, including: guiding the self-reflection module to rewrite the CoT planning, adding a logical verification step, or adopting a counterfactual reasoning verification mechanism to improve the decision confidence. The instruction is directly fed back to the prompt word structure of the large language model or the decision architecture of the Agent.

[0078] The above optimization instruction will be input into the downstream optimization module of the self-evolution intelligent agent system as a highest priority feedback signal, ensuring that the system can self-correct and ability leap in the continuous evolution with the highest efficiency and verifiable accuracy, thereby realizing the nonlinear, stepwise growth of the task success rate. It can be understood that under the self-evolution intelligent agent system, the intelligent agent does not rely on artificial intervention, but continuously improves its ability, strategy, knowledge or structure through its own experience, reflection, trial and error or generation of new data. Therefore, it is more dependent on the accuracy of fault analysis, and in the related art, because it is impossible to distinguish whether the logical loophole in the reasoning planning or the defect of the knowledge stored in the knowledge graph, the system may waste a large number of iteration times and computing resources to optimize the correct reasoning process, while ignoring the outdated knowledge that really needs to be updated. By mapping the uncertainty, knowledge defect and other cognitive factors in the intelligent agent decision to a single, computable physical quantity, high-precision, verifiable causal attribution is achieved.

[0079] By introducing the physical impedance model, the application transfers the wrong decision weight from the fuzzy semantic probability to the computable physical conservation quantity, realizes accurate multi-granularity causal attribution, and can realize quantitative and accurate attribution of the root cause of failure of complex tasks in the self-evolution intelligent agent system, solving the core problems of qualitative fault attribution, inability to separate planning errors and knowledge defects, and lack of unified quantitative measurement in the related art. By introducing the impedance (Impedance, Z) concept in circuit theory, the decision process of the intelligent agent is formalized as a hierarchical computing circuit. By quantitatively calculating the resistance degree of each step of decision to the information flow in the circuit, the contribution degree of the reasoning planning error and the knowledge timeliness defect is accurately distinguished, and then a high-precision mental optimization instruction is generated to enable the self-evolution intelligent agent to realize efficient and reliable continuous learning. By converting the abstract cognitive failure into a computable physical quantity, the reliability and accuracy of fault attribution are greatly improved, and the efficiency of subsequent optimization is improved.

[0080] As can be seen from the above, in the embodiment, a traceable log generated by an agent performing a target task is acquired, and a circuit topology is constructed based on the traceable log; one node in the circuit topology corresponds to one decision in a decision sequence of the target task, and a connection relationship between nodes in the circuit topology represents a logical relationship between different decisions; according to a decision confidence, an action complexity and a knowledge defect degree corresponding to each node in the circuit topology, a local computing impedance corresponding to each node is determined by using a computing impedance model; a target node is screened out according to the local computing impedance, an agent fault reason is determined according to an impedance composition of the target node, and an optimization instruction for agent updating is generated. As can be seen from the above, by mapping a task execution process of an agent to a circuit topology and using an impedance model, a decision confidence, an action complexity and a knowledge defect degree of the agent are quantified as an impedance in the circuit, so that an agent fault reason is determined by analyzing the impedance, and an optimization instruction for agent updating is generated according to the agent fault reason, thereby improving the accuracy of agent fault analysis and the efficiency and reliability of agent optimization.

[0081] Correspondingly, the embodiment of the application further discloses an agent fault attribution device based on a computing impedance model, as shown in Figure 2 The device comprises:

[0082] A circuit topology construction module 11 is configured to acquire a traceable log generated by an agent performing a target task, and construct a circuit topology based on the traceable log; one node in the circuit topology corresponds to one decision in a decision sequence of the target task, and a connection relationship between nodes in the circuit topology represents a logical relationship between different decisions;

[0083] A local computing impedance calculation module 12 is configured to determine a local computing impedance corresponding to each node by using a computing impedance model according to a decision confidence, an action complexity and a knowledge defect degree corresponding to each node in the circuit topology;

[0084] A fault reason determination module 13 is configured to screen out a target node according to the local computing impedance, determine an agent fault reason according to an impedance composition of the target node, and generate an optimization instruction for agent updating.

[0085] As can be seen from the above, in the embodiment, a traceable log generated by an agent performing a target task is acquired, and a circuit topology is constructed based on the traceable log; one node in the circuit topology corresponds to one decision in a decision sequence of the target task, and a connection relationship between nodes in the circuit topology represents a logical relationship between different decisions; according to a decision confidence, an action complexity and a knowledge defect degree corresponding to each node in the circuit topology, a local computing impedance corresponding to each node is determined by using a computing impedance model; a target node is screened out according to the local computing impedance, a fault reason of the agent is determined according to an impedance composition of the target node, and an optimization instruction for updating the agent is generated. As can be seen from the above, by mapping a task execution process of the agent to the circuit topology and using the impedance model, the decision confidence, the action complexity and the knowledge defect degree involved in the decision of the agent are quantified as impedances in the circuit, so that the fault reason of the agent is determined by analyzing the impedances, and the optimization instruction for updating the agent is generated according to the fault reason, thereby improving the accuracy of agent fault analysis and the efficiency and reliability of agent optimization.

[0086] In some specific embodiments, the local computing impedance calculation module 12 can specifically include:

[0087] A local computing impedance calculation unit is configured to obtain, by weighting calculation, a local computing impedance corresponding to a current node based on a decision confidence corresponding to the current node, an entropy of an action sequence performed by the current node, and a difference degree between new knowledge retrieved by the current node and old knowledge in the agent, and weights corresponding to each of the above.

[0088] The decision confidence and the local computing impedance are negatively correlated, the entropy and the local computing impedance are positively correlated, and the difference degree and the local computing impedance are positively correlated.

[0089] In some specific embodiments, the agent fault attribution device can specifically include:

[0090] A reference total computing impedance calculation unit is configured to calculate a reference total computing impedance corresponding to the circuit topology after determining the local computing impedance corresponding to each node.

[0091] An actual total computing impedance unit is configured to calculate an actual total computing impedance corresponding to the circuit topology by circuit analysis according to the local computing impedance of each node and the circuit topology.

[0092] A weight updating unit is configured to update weights of the decision confidence, the entropy and the difference degree according to a difference degree between the reference total computing impedance and the actual total computing impedance.

[0093] In some embodiments, the reference total computing impedance calculation unit is configured to determine a target parameter as a task success score according to a task type of the target task; determine a task failure severity based on the task success score; determine a computing throughput according to valid information in a process of executing the task by the agent; and calculate a total computing impedance corresponding to the task by taking the task failure severity as an electric potential difference and taking the computing throughput as an effective current.

[0094] In some embodiments, the agent failure attribution device can include an effective information statistics unit configured to perform syntax analysis on content generated by the agent before determining the computing throughput according to the valid information in the process of executing the task by the agent, and take content meeting syntax validity as the valid information; perform relevance threshold checking on knowledge searched by the agent, and take knowledge passing the checking as the valid information; and take a number of conflict-free delivered messages as the valid information.

[0095] In some embodiments, the failure cause determination module 13 can include:

[0096] A target node screening unit is configured to take a node corresponding to the maximum local computing impedance as a target node, or take N nodes with the highest local computing impedance values as target nodes, or take a node with a local computing impedance greater than a preset local computing impedance threshold as a target node.

[0097] In some embodiments, the traceable log includes cognitive data and execution data, the cognitive data includes a thinking chain and a tool call sequence, and the execution data includes a stack trace and a running-related error code.

[0098] In some embodiments, the failure cause determination module 13 can include:

[0099] A first optimization unit is configured to generate a first optimization instruction if a contribution degree of a knowledge defect degree of the target node to the local computing impedance is greater than a contribution degree threshold; and determine a new weight of a target old knowledge point based on a weight of the target old knowledge point and a forgetting rate parameter according to the first optimization instruction, or perform isolation processing on the target old knowledge point; the target old knowledge point is a knowledge point related to the knowledge defect degree in an agent knowledge graph.

[0100] The second optimization unit is configured to generate a second optimization instruction if the contribution degree of the action complexity of the target node to the local computing impedance is greater than a contribution degree threshold; and according to the second optimization instruction, the action sequence performed by the target node is disassembled into an atomic operation sequence, and / or a parameter constraint mode of a tool is started, and / or an application program interface is called in a descending order of priority; wherein the higher the priority is, the more single the function of the application program interface is, and the higher the determinacy is.

[0101] The third optimization unit is configured to generate a third optimization instruction if the contribution degree of the decision confidence of the target node to the local computing impedance is greater than a contribution degree threshold; and according to the third optimization instruction, a thinking chain planning framework is updated, and / or a logic verification rule for a decision in a decision sequence is added, and / or a counterfactual reasoning verification mechanism is increased.

[0102] Further, the embodiment of the present application also discloses an electronic device, which comprises a memory for saving a computer program and a processor for executing the computer program to realize the intelligent agent fault attribution method based on the computing impedance model. Figure 3 It should be noted that the contents shown in the drawings cannot be considered as any limitation on the use range of the present application.

[0103] Figure 3 A structural schematic diagram of an electronic device 20 provided by the embodiment of the present application is shown. The electronic device 20 specifically can comprise at least one processor 21, at least one memory 22, a power supply 23, a communication interface 24, an input / output interface 25 and a communication bus 26. The memory 22 is used to store a computer program, and the computer program is loaded and executed by the processor 21 to realize the related steps in the intelligent agent fault attribution method based on the computing impedance model disclosed in any of the foregoing embodiments.

[0104] In the embodiment, the power supply 23 is used to provide working voltage for each hardware device on the electronic device 20; the communication interface 24 can create a data transmission channel between the electronic device 20 and external devices, and the communication protocol followed by the communication interface 24 can be any communication protocol applicable to the technical solution of the present application, which is not limited specifically herein; the input / output interface 25 is used to obtain external input data or output data to the outside, and the specific interface type can be selected according to the specific application needs, which is not limited specifically herein.

[0105] In addition, the memory 22 as a carrier of resource storage can be a read-only memory, a random access memory, a magnetic disk or an optical disk, etc., and the resources stored thereon include an operating system 221, a computer program 222 and data 223 including traceable logs, etc., and the storage mode can be temporary storage or permanent storage.

[0106] The operating system 221 is used to manage and control each hardware device on the electronic device 20 and the computer program 222, so as to realize the operation and processing of the processor 21 on the massive data 223 in the memory 22, which can be Windows Server, Netware, Unix, Linux, etc. In addition to the computer program capable of completing the intelligent agent fault attribution method based on the calculation impedance model executed by the electronic device 20 disclosed in any one of the foregoing embodiments, the computer program 222 can further include a computer program capable of completing other specific work.

[0107] Further, the embodiment of the present application further discloses a computer storage medium, which stores computer executable instructions. When the computer executable instructions are loaded and executed by a processor, the steps of the intelligent agent fault attribution method based on the calculation impedance model disclosed in any one of the foregoing embodiments are realized.

[0108] Further, the embodiment of the present application further discloses a computer program product, which includes a computer program. When the computer program is executed by a processor, the steps of the intelligent agent fault attribution method based on the calculation impedance model disclosed in any one of the foregoing embodiments are realized.

[0109] The embodiments in the specification are described in a progressive manner, and each embodiment focuses on the difference from other embodiments. The same or similar parts of each embodiment can be referred to each other. For the device disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple, and the related parts can be referred to the method part.

[0110] The steps of the method or algorithm described in combination with the embodiments disclosed herein can be directly implemented by hardware, a software module executed by a processor, or a combination of the two. The software module can be placed in a random access memory (RAM), a memory, a read-only memory (ROM), an electrically programmable ROM, an electrically erasable programmable ROM, a register, a hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the technical field.

[0111] Finally, it needs to be pointed out that, in this article, the relationship terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between the entities or operations. Moreover, the term "includes", "contains" or any other variant thereof is intended to cover non-exclusive inclusion, so that the process, method, article or device including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such process, method, article or device. Without more limitations, the element defined by the statement "includes a" does not exclude the presence of other identical elements in the process, method, article or device including the element.

[0112] The above describes in detail the intelligent agent fault attribution method, device, equipment and storage medium based on a computing impedance model provided by the present application. The principles and implementation manners of the present application are described by applying specific examples in this article. The above description of the embodiments is only used to help understand the method of the present application and its core idea. Meanwhile, for those skilled in the art, according to the idea of the present application, the specific implementation manners and application ranges will be changed. In view of the above, the content of the specification should not be understood as a limitation of the present application.

Claims

1. A fault attribution method for intelligent agents based on a computational impedance model, characterized in that, The method comprises the following steps: acquiring a traceable log generated by an agent performing a target task, and constructing a circuit topology based on the traceable log; one node in the circuit topology corresponds to one decision in a decision sequence of the target task, and a connection relationship between nodes in the circuit topology represents a logical relationship between different decisions; determining a local computing impedance corresponding to each node in the circuit topology according to a decision confidence, an action complexity and a knowledge defect degree corresponding to each node in the circuit topology, and using a computing impedance model; screening a target node according to the local computing impedance, determining a fault reason of the agent according to an impedance composition of the target node, and generating an optimization instruction for updating the agent; wherein, determining the local computing impedance corresponding to each node according to the decision confidence, the action complexity and the knowledge defect degree corresponding to each node in the circuit topology, and using the computing impedance model, comprises: obtaining the local computing impedance corresponding to the current node by weighted calculation based on the decision confidence corresponding to the current node, the entropy of the action sequence performed by the current node, and the difference degree between the new knowledge retrieved by the current node and the old knowledge in the agent, and the weights corresponding to each item; the decision confidence and the local computing impedance are negatively correlated, the entropy and the local computing impedance are positively correlated, and the difference degree and the local computing impedance are positively correlated.

2. The computational impedance model based agent fault attribution method of claim 1, wherein, After determining the local computing impedance corresponding to each node, the method further comprises: calculating a reference total computing impedance corresponding to the circuit topology; calculating an actual total computing impedance corresponding to the circuit topology by circuit analysis according to the local computing impedance of each node and the circuit topology; updating the weights of the decision confidence, the entropy and the difference degree according to the difference degree between the reference total computing impedance and the actual total computing impedance.

3. The computational impedance model based agent fault attribution method of claim 2, wherein, The method of calculating the reference total computing impedance corresponding to the circuit topology comprises: determining a target parameter as a task success score according to the task type of the target task; determining a task failure severity based on the task success score, and determining a computing throughput according to effective information in the agent task execution process; calculating a total computing impedance corresponding to the task by taking the task failure severity as a potential difference and taking the computing throughput as an effective current.

4. The method of claim 3, wherein, Before determining the computing throughput according to the effective information in the agent task execution process, the method further comprises: performing syntax analysis on the content generated by the agent, and taking the content meeting the syntax validity as the effective information; performing relevance threshold verification on the knowledge retrieved by the agent, and taking the knowledge passing the verification as the effective information; taking the number of conflict-free delivered messages as the effective information.

5. The method of claim 1, wherein, The method of screening the target node according to the local computing impedance comprises: taking the node corresponding to the maximum local computing impedance as the target node, or taking the first N nodes in the local computing impedance value sorting as the target node, or taking the node with a local computing impedance greater than a preset local computing impedance threshold as the target node.

6. The computational impedance model based agent fault attribution method of claim 1, wherein, The traceable log comprises cognitive data and execution data, the cognitive data comprises a thinking chain and a tool calling sequence, and the execution data comprises a stack trajectory and an error code related to running.

7. The computational impedance model based agent fault attribution method according to any one of claims 1 to 6, characterized in that, According to the impedance composition of the target node, the agent failure cause is determined, and corresponding optimization instructions for agent updating are generated, including: If the contribution degree of the knowledge defect degree of the target node to the local computing impedance is greater than the contribution degree threshold, a first optimization instruction is generated; According to the first optimization instruction, the new weight of the target old knowledge point is determined based on the weight of the target old knowledge point and the forgetting rate parameter, or the target old knowledge point is isolated; the target old knowledge point is a knowledge point related to the knowledge defect degree in the agent knowledge graph; If the contribution degree of the action complexity of the target node to the local computing impedance is greater than the contribution degree threshold, a second optimization instruction is generated; According to the second optimization instruction, the action sequence executed by the target node is disassembled into an atomic operation sequence, and / or the parameter constraint mode of tool calling is started, and / or the application program interface is called in order of priority from high to low; the higher the priority, the more single the function of calling the application program interface, and the higher the priority, the higher the certainty; If the contribution degree of the decision confidence of the target node to the local computing impedance is greater than the contribution degree threshold, a third optimization instruction is generated; According to the third optimization instruction, the thought chain planning framework is updated, and / or the logic verification rule for the decision in the decision sequence is added, and / or the counterfactual reasoning verification mechanism is increased.

8. A fault attribution device for intelligent agents based on a computational impedance model, characterized in that, Including: A circuit topology construction module is configured to obtain a traceable log generated by an agent executing a target task, and construct a circuit topology based on the traceable log; One node in the circuit topology corresponds to one decision in a decision sequence of a target task, and the connection relationship between nodes in the circuit topology represents the logical relationship between different decisions; A local computing impedance calculation module is configured to determine the local computing impedance of each node in the circuit topology according to the decision confidence, action complexity and knowledge defect degree corresponding to each node in the circuit topology by using a computing impedance model; A failure cause determination module is configured to filter out a target node according to the local computing impedance, determine the agent failure cause according to the impedance composition of the target node, and generate corresponding optimization instructions for agent updating; The local computing impedance calculation module is configured to obtain the local computing impedance corresponding to the current node by weighting calculation based on the decision confidence corresponding to the current node, the entropy of the action sequence executed by the current node, and the difference degree between the new knowledge retrieved by the current node and the old knowledge in the agent, as well as the weights of each item; the decision confidence and the local computing impedance are negatively correlated, the entropy and the local computing impedance are positively correlated, and the difference degree and the local computing impedance are positively correlated.

9. A computer-readable storage medium, characterized in that, A computer program is stored in the computer program; when the computer program is executed by the processor, the agent failure attribution method based on the computing impedance model is realized. The computer program is stored in the computer program; when the computer program is executed by the processor, the agent failure attribution method based on the computing impedance model is realized.

Images