A method, system, and device for training management of an agent

Abstract

The application provides a kind of training management method, system and equipment of intelligent agent, belong to intelligent agent technical field, specifically include: with the interaction demand type between virtual user and intelligent agent as foundation, and combine the change situation of interaction user data in different time periods, the determination of available interaction period between virtual object and intelligent agent, according to available interaction period data, the determination of interactive data analysis processing strategy between interaction user and intelligent agent, according to the analysis processing result of interactive data of virtual object in available interaction period, the interactive data between interaction user and intelligent agent, determine the interactive management strategy of virtual object in the period outside available interaction period, improve the efficiency of intelligent agent update processing, also reduce the influence to normal business process.

Description

Technical Field

[0001] This invention belongs to the field of intelligent agent technology, and in particular relates to a training and management method, system and device for intelligent agents. Background Technology

[0002] Existing customer service systems generally suffer from cold start difficulties. When applied to new fields, the lack of relevant knowledge may prevent the agent from effectively solving user problems. Therefore, how to achieve cold start for the agent has become an urgent technical problem to be solved.

[0003] To address the aforementioned technical problems, existing solutions often employ transfer learning for agent training. However, these solutions suffer from the following technical issues:

[0004] By constructing virtual users through transfer learning and enabling them to interact with intelligent agents, the efficiency of identifying and processing the reliability of agent interactions can be improved, and targeted training and updating of the agent can be performed. However, since the number of interaction ports of the agent is limited, determining the available interaction time of virtual users and the corresponding interaction management strategies between virtual users and intelligent agents, while improving the processing efficiency of agent training and updating without affecting normal user use, has become an urgent technical problem to be solved.

[0005] Therefore, there is an urgent need for a training and management method, system, and device for intelligent agents. Summary of the Invention

[0006] To achieve the objectives of this invention, the following technical solution is adopted:

[0007] Specifically, this application provides a training management method for intelligent agents, which includes:

[0008] S1 constructs a virtual user based on the interaction data of the intelligent agent. Based on the interaction data between the intelligent agent and the interactive user, it determines the matching situation of the interaction needs between the intelligent agent and the interactive user. Based on the matching situation and the interactive user data in different time periods, if it is determined that the interaction need type between the virtual user and the intelligent agent does not belong to the preset need type, it proceeds to the next step.

[0009] S2 determines the available interaction time periods between the virtual user and the intelligent agent based on the interaction demand type between the virtual user and the intelligent agent, and in combination with the changes in the interaction user data in different time periods. Based on the available interaction time period data, the analysis and processing strategy for the interaction data between the interactive user and the intelligent agent is determined.

[0010] S3 determines the interaction management strategy for the virtual object in time periods outside the available interaction period based on the analysis and processing results of the interaction data of the virtual object and the interaction data between the interactive user and the intelligent agent during the available interaction period.

[0011] The beneficial effects of this invention are as follows:

[0012] The available interaction time periods between virtual objects and intelligent agents are determined by considering the types of interaction needs between virtual users and intelligent agents and the changes in user data at different times. This approach takes into account both the need for intelligent agents to use virtual users for training and update processing and the drastic changes in the number of interactive users at different times. As a result, the available interaction time periods can be determined without affecting normal user use, thus ensuring the reliability of the intelligent agent's update processing.

[0013] Based on the analysis and processing results of the interaction data of virtual objects and the interaction data between interactive users and agents during the available interaction periods, the interaction management strategy of virtual objects in the time periods outside the available interaction periods is determined. This takes into account both the interaction deviation of virtual objects in different available interaction periods and the recognition deviation of agents when interacting with interactive users. Thus, the interaction management strategy in other time periods is determined based on the interaction deviation and recognition deviation, which further improves the efficiency of agent training and update processing.

[0014] Furthermore, the virtual user is a virtual user constructed based on the interaction data between the intelligent agent and the interactive user. The virtual user interacts with the intelligent agent to identify the intelligent agent's recognition deviation of the interaction data.

[0015] It should be noted that the interactive user is the user who interacts with the intelligent agent to exchange data.

[0016] Furthermore, the matching of the interaction needs between the intelligent agent and the interactive user is determined based on the number of times the intelligent agent and the interactive user are transferred to the human system during the interaction process.

[0017] Furthermore, the interactive user data includes the number of interactive users during the time period on different dates.

[0018] Furthermore, determining that the interaction request type between the virtual user and the intelligent agent does not belong to the preset request type specifically includes:

[0019] Based on the matching results, the number of times the agent and the user are switched to the human system during the interaction process is determined and used as the switching count;

[0020] Based on the interactive user data of the agent in different time periods, the total duration of idle ports of the agent in different time periods is determined and used as the idle duration;

[0021] Based on the number of switching times and the idle time in different time periods, it is determined whether the interaction demand type between the virtual user and the intelligent agent belongs to the preset demand type.

[0022] Furthermore, the method for determining the interaction management strategy of the virtual object during periods outside the available interaction time is as follows:

[0023] Based on the interaction data of the virtual objects during the available interaction period, determine the number of interactions between the virtual objects and the agent on different dates;

[0024] Based on the analysis and processing results of the interaction data between the interactive user and the intelligent agent, the number of interactions in which the semantic understanding of the interaction data between the interactive user and the intelligent agent is determined, and this number is taken as the number of understanding deviations.

[0025] Based on the number of interactions between the virtual object and the agent on different dates, the number of comprehension deviations, and the analysis and processing strategy, the interaction management strategy for the virtual object in time periods outside the available interaction period is determined.

[0026] In a second aspect, the present invention provides a computer system comprising: a memory and a processor connected in communication, and a computer program stored in the memory and capable of running on the processor, wherein the processor executes the above-described method for training and managing an intelligent agent when running the computer program.

[0027] Thirdly, this application provides a training management device for an intelligent agent, employing the aforementioned training management method for an intelligent agent, specifically including:

[0028] The module includes an interaction time segmentation module, an analysis strategy determination module, and an interaction management module.

[0029] The interaction time segmentation module is responsible for determining whether the interaction request type between the virtual user and the intelligent agent belongs to the preset request type.

[0030] The analysis strategy determination module is responsible for determining the analysis and processing strategy for the interaction data between the interactive user and the intelligent agent.

[0031] The interaction management module is responsible for determining the interaction management strategy for virtual objects during periods outside of the available interaction time slots.

[0032] Other features and advantages will be set forth in the following description, and the objects and other advantages of the invention are realized and obtained through the structures particularly pointed out in the description and the drawings.

[0033] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0034] The above and other features and advantages of the present invention will become more apparent from a detailed description of exemplary embodiments thereof with reference to the accompanying drawings.

[0035] Figure 1 This is a flowchart of a training and management method for intelligent agents;

[0036] Figure 2 This is a flowchart for determining whether the interaction request type between virtual users and intelligent agents does not belong to the preset request type;

[0037] Figure 3 This is a flowchart illustrating the method for determining the available interaction time between virtual objects and intelligent agents;

[0038] Figure 4 This is a flowchart illustrating the method for determining the analysis and processing strategy of interaction data between users and intelligent agents. Detailed Implementation

[0039] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this specification, and not all embodiments. Based on the embodiments of this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this specification.

[0040] Example 1

[0041] like Figure 1 As shown, this application provides a training management method for intelligent agents, specifically including:

[0042] S1 constructs a virtual user based on the interaction data of the intelligent agent. Based on the interaction data between the intelligent agent and the interactive user, it determines the matching situation of the interaction needs between the intelligent agent and the interactive user. Based on the matching situation and the interactive user data in different time periods, if it is determined that the interaction need type between the virtual user and the intelligent agent does not belong to the preset need type, it proceeds to the next step.

[0043] Furthermore, the virtual user is a virtual user constructed based on the interaction data between the intelligent agent and the interactive user. The virtual user interacts with the intelligent agent to identify the intelligent agent's recognition deviation of the interaction data.

[0044] It should be noted that the interactive user is the user who interacts with the intelligent agent to exchange data.

[0045] Furthermore, the matching of the interaction needs between the intelligent agent and the interactive user is determined based on the number of times the intelligent agent and the interactive user are transferred to the human system during the interaction process.

[0046] Furthermore, the interactive user data includes the number of interactive users during the time period on different dates.

[0047] Specifically, such as Figure 2 As shown, determining that the interaction request type between the virtual user and the intelligent agent does not belong to the preset request type specifically includes:

[0048] Identification bias: refers to situations where the intelligent agent misunderstands, misguides, or is unable to handle user requests. The goal of this embodiment is to identify these biases, rather than to correct them during testing.

[0049] Normal users (interactive users): Real consumers on the platform whose interactions with the intelligent agents generate traffic in real time, requiring assurance of their service experience and response speed.

[0050] Virtual users: Simulated users generated by the system, whose sole purpose is to engage in test conversations with the intelligent agent to expose potential recognition biases. Interactions with virtual users do not generate actual orders or after-sales service tickets.

[0051] The core challenge is that allowing virtual users to interact with the intelligent agent without restriction could consume system resources, causing regular user requests to queue or respond more slowly. Therefore, testing must intelligently schedule virtual users based on the agent's idle status.

[0052] S11 Based on the matching situation, determine the number of times the intelligent agent and the interactive user are transferred to the human system during the interaction process, and use this as the number of switching times;

[0053] The core metric is the number of switching attempts. Switching attempts refer to the number of times a real user (interacting user) is actively or automatically transferred to human customer service during a single conversation with the agent due to unresolved issues or a poor experience. Only one switching attempt is recorded per conversation (regardless of the number of transfer attempts in between).

[0054] The matching status between the intelligent agent and the interactive user's interaction needs: This is a qualitative conclusion, quantified by the core indicator of the number of switching attempts. A high number of switching attempts indicates a poor matching status.

[0055] S12 determines the total duration of idle ports of the intelligent agent in different time periods based on the interactive user data of the intelligent agent in different time periods, and uses it as the idle time.

[0056] Interactive user data (time period dimension): refers to the number of unique real users who interact with the agent within a specific time period (such as "14:00-14:30" every day).

[0057] Idle Port: A channel in the intelligent agent system's preset concurrent processing channels that is not occupied by real user sessions.

[0058] Idle time: The total cumulative time (in minutes) during which the system has at least one idle port within a specific time period. For example, if the system is in an "idle port available" state for 10 minutes out of the 30 minutes from "14:00 to 14:30", then the idle time is 10 minutes.

[0059] The platform analyzes the agent's load monitoring data for different time periods (e.g., "14:00-14:30") over the past 7 days. Example data: During the "14:00-14:30" period, the average number of interactive users over the 7 days was 800 per time period, and the system's maximum concurrency was 1000. Monitoring logs show that there were an average of 15 minutes of idle ports during this period. Calculation result: Idle time = 15 minutes / time period

[0060] Initiating testing solely based on a high "switch count" could increase system load during peak business periods. Introducing "idle time" assesses the feasibility of testing from the perspective of system resource availability, ensuring that testing does not interfere with normal services and providing a basis for resource safety boundaries for subsequent decisions.

[0061] S13 determines whether the interaction request type between the virtual user and the intelligent agent belongs to the preset request type based on the number of switching times and the idle time in different time periods.

[0062] It is understood that, based on the number of switching attempts and the idle time in different time periods, it is determined whether the interaction request type between the virtual user and the intelligent agent belongs to a preset request type, specifically including:

[0063] S131 Based on the number of switching, determine the proportion of the number of switching in the number of interactions between the agent and the user, and use it as the proportion of the number of switching. Determine whether the proportion of the number of switching is greater than a preset threshold for the proportion of the number of switching. If yes, determine that the interaction requirement type between the virtual user and the agent belongs to a preset requirement type. If no, proceed to the next step.

[0064] In the above steps, the percentage of switching times is calculated and a preliminary judgment is made. The number of interactions between the agent and the user is the total number of times the user initiates a dialogue with the agent for personal needs within the statistical period. The percentage of switching times is calculated as follows: Percentage of switching times = (Number of switching times) / (Number of interactions). The preset threshold for the percentage of switching times is the critical value of the key performance indicator (KPI) set by the platform, such as 25%.

[0065] Given: Number of interactions = 1000 times, number of switches = 300 times, calculate: percentage of switches = 300 / 1000 = 30%, judge: 30% > 25% (preset threshold), since the percentage of switches is greater than the threshold, directly enter the "yes" path of branch S131.

[0066] The system determines that the interaction needs between virtual users and intelligent agents belong to preset need types. By setting clear quantitative thresholds, it achieves automated and rapid identification of high-risk need types without having to go through more complex subsequent judgments, thus improving decision-making efficiency.

[0067] S132 determines whether the percentage of switching times is less than a preset percentage threshold. If yes, it determines that the interaction requirement type between the virtual user and the intelligent agent belongs to the second type of requirement and does not belong to the preset requirement type. If no, it proceeds to the next step.

[0068] S133 Based on the idle time of the intelligent agent in different time periods, determine the idle time periods in the time periods, and determine whether there are idle time periods in different dates. If so, determine that the interaction demand type between the virtual user and the intelligent agent belongs to the preset demand type. If not, proceed to the next step.

[0069] S134 determines whether the interaction request type between the virtual user and the intelligent agent needs to be preset based on the average number of idle time periods in different dates.

[0070] It is understandable that when the average number of idle time slots on different dates is greater than the preset threshold for the number of idle time slots, the interaction demand type between the virtual user and the intelligent agent is determined to belong to the preset demand type. Otherwise, the interaction demand type between the virtual user and the intelligent agent is determined not to belong to the preset demand type, but to belong to a type of demand.

[0071] In another possible embodiment, assuming that when the data changes: number of interactions: 5000 times, number of switches: 600 times, percentage of switches = 600 / 5000 = 12%, preset threshold for percentage of switches = 25%, preset threshold for percentage (lower limit) = 5%, preset threshold for number of idle time periods = 4 periods / day.

[0072] S132 makes a judgment: 12% < 25% but > 5%, which does not meet the "less than the preset percentage threshold" requirement, so proceed to the next step.

[0073] S133 Decision (requires determining "idle time period"): Idle time period: A time period within which the idle duration exceeds a preset standard (e.g., idle duration > 20 minutes). The system analyzes high-incidence consultation times (e.g., 10:00-10:30) and checks whether these times have "idle duration > 20 minutes" every day for the past 7 days. It is found that the business volume is huge, and the "idle time period" condition is only met in the early morning (e.g., 02:00-06:00) (e.g., 2:00-2:30). However, not all relevant time periods on peak daytime dates meet this condition. Therefore, the decision is "No," proceeding to S134.

[0074] S134 determines the number of idle time slots in different dates: This refers to the number of time slots that are identified as "idle time slots" each day within the statistical period (e.g., the past 7 days). The preset threshold for the number of idle time slots is, for example, 4 slots / day.

[0075] Implementation: Calculate the average number of times the "idle time" condition is met during the peak periods of this demand each day over the past 7 days. Assuming an average of 2 times / day, if 2 times / day < 4 times / day (preset threshold), the interaction demand type between the virtual user and the intelligent agent for this demand type (ordinary logistics query) does not belong to the preset demand type, but is a separate demand type.

[0076] For demands with a switching frequency ratio in the "middle range" (neither high nor low), a more refined secondary judgment is needed, taking into account both the sustainability of system resource availability and the dispersion of changes in user demand. This avoids misclassifying demands with slightly higher switching rates due to temporary peak periods or frequent changes in user demand as "preset demand types," thus achieving hierarchical and categorized management.

[0077] It should be noted that if the interaction request type between the virtual user and the intelligent agent belongs to a preset request type, then the number of interactions between the virtual user and the intelligent agent will be determined according to a preset proportion of the number of idle ports, as long as there are idle ports in different time periods.

[0078] Specifically, the virtual user scheduling is preset to the required type, the idle port is the processing channel in the agent system that is not occupied by real user sessions at a certain moment, and the preset ratio is a coefficient between 0 and 1 (e.g., 0.2) used to control the upper limit of resources occupied by virtual user tests.

[0079] Implementation (following S131 determination): When a system has been determined to be of "preset demand type": The system continuously monitors the port occupancy of the intelligent agent in each time period. Whenever an idle port is detected in a certain time period (e.g., there are currently 200 idle ports), a virtual user test is initiated.

[0080] Calculate the number of virtual user interactions: Number of virtual users = Number of idle ports × Preset ratio. For example, if the preset ratio is 0.2, then this time 200 × 0.2 = 40 virtual users will be launched. These 40 virtual users will simulate various complex scenarios (such as product A participating in a 300-40 discount, product B participating in a 200-20 discount, and the user having a category coupon) to interact with the agent and record all responses.

[0081] By employing a dual mechanism of "triggering when idle ports exist" and "proportional utilization," the virtual test is ensured to be conducted strictly within the redundancy of system resources, achieving "zero interference" with normal business operations.

[0082] It should be noted that the preset ratio is between 0 and 1.

[0083] S2 determines the available interaction time periods between the virtual user and the intelligent agent based on the interaction demand type between the virtual user and the intelligent agent, and in combination with the changes in the interaction user data in different time periods. Based on the available interaction time period data, the analysis and processing strategy for the interaction data between the interactive user and the intelligent agent is determined.

[0084] Specifically, such as Figure 3 As shown, the method for determining the available interaction time period between the virtual object and the intelligent agent is as follows:

[0085] S31 uses interactive user data in different time periods to determine the time periods, and the idle time periods in different dates are defined as the dates in the time periods where there are idle time periods;

[0086] In the above steps, the system determines the idle dates and counts the number of days in the past 14 days that are considered "idle periods". All dates are considered idle dates.

[0087] S32 determines the number of idle time slots in the time period based on the changes in interactive user data during the time period;

[0088] In the above steps, the number of idle time slots is determined. The system confirms that on each idle date, there is only one consecutive idle time slot.

[0089] S33 determines whether the time period is an available interaction period between the virtual object and the intelligent agent based on the idle date data in the time period, the number of idle time periods in the time period for different dates, and the interaction requirement type.

[0090] It is understood that, based on the idle date data within the time period, the number of idle time periods within the time period for different dates, and the type of interaction request, determining whether the time period is an available interaction period between the virtual object and the intelligent agent specifically includes:

[0091] S331 Based on the number of idle time periods in the time period on different dates, determine the average number of idle time periods in the time period on different dates, and determine whether the average number of idle time periods is less than a preset idle time period threshold. If yes, proceed to step S332. If no, then because the busy state often changes, using virtual users for interactive processing in the time period will inevitably affect the normal interaction between the intelligent agent and the interactive user. Therefore, it is determined that the time period does not belong to the available interaction time period between the virtual object and the intelligent agent.

[0092] In the above steps, the average number of idle time slots is determined. The average number of idle time slots = 1 slot / day. The preset threshold for the number of idle time slots is: assuming the platform is set to 2 slots / day (this threshold is relatively high and is used to filter out extremely stable idle time slots). Judgment: 1<2 → Satisfies the "yes" condition. Decision: Proceed to step S332.

[0093] If the average value of 1 is less than the threshold of 2, it indicates that although the period is idle every day, the idle pattern is very stable (the entire period is a complete idle block). If the average value is close to or exceeds 2, it means that the idle-busy period may switch frequently and fluctuate greatly, making it unsuitable for starting a long-term full test, and should be directly excluded (follow the S331 "No" path).

[0094] S332 Obtain the free dates, and determine whether the proportion of the free dates in all dates is greater than the preset proportion of the number of free dates. If yes, proceed to the next step; otherwise, proceed to step S334.

[0095] In the above steps, determine the percentage of free days and calculate: percentage of free days = 100%. The preset percentage of free days is assumed to be 80%. Determine: 100% > 80% → the "yes" condition is met. Decision: Proceed to step S333.

[0096] A 100% idle rate indicates that this period is historically absolutely reliable and is a perfect candidate window for conducting full-scale testing.

[0097] S333 determines whether the average duration of the idle period on different idle dates is less than a preset duration threshold. If yes, proceed to the next step; otherwise, determine that the period belongs to the available interaction period between the virtual object and the intelligent agent.

[0098] In the above steps, the idle time of the idle date is determined and calculated as follows: average idle time ≈ 29.1 minutes. The preset time threshold is assumed to be 20 minutes. The judgment is: 29.1 minutes > 20 minutes → the condition "no" is met.

[0099] Final decision (based on S333 logic): Determine that this period belongs to the available interaction period between virtual objects and intelligent agents. The average idle time is close to the length of the entire period (30 minutes), which means that there is a very sufficient and continuous time to execute large-scale, uninterrupted virtual user tests. The test tasks can be fully carried out without being interrupted by system busy.

[0100] S334 determines whether the proportion of the idle dates in all dates is less than a preset date quantity proportion threshold. If yes, it determines that the time period does not belong to the available interaction time period between the virtual object and the intelligent agent. If no, it proceeds to the next step.

[0101] In the above steps, it is determined whether the percentage is less than a lower threshold (a preset date quantity percentage threshold, for example, set to 20%). 28.6%>20% → does not meet the "less than" condition, proceed to S335.

[0102] S335 determines whether the interaction request type is a type of request. If yes, it determines that the time period belongs to the available interaction time period between the virtual object and the intelligent agent. If no, it determines that the time period does not belong to the available interaction time period between the virtual object and the intelligent agent.

[0103] To determine whether the interaction requirement type in this test is a "Type 1 requirement," this is a "full-scale test," which is a typical test task that is neither "preset" nor "Type 1." Based on the preceding logic, it is classified as "not belonging to the preset requirement type" and belongs to "Type 2 requirement type." Its identification bias is relatively low; therefore, the judgment result is "no," and this time period is determined not to be a usable interaction period.

[0104] Specifically, such as Figure 4 As shown, the method for determining the analysis and processing strategy of the interaction data between the user and the intelligent agent is as follows:

[0105] S41 determines the number of available interaction time periods based on the available interaction time period data;

[0106] Specifically, the system automatically counts the total number of available interaction periods within the analysis period, determining the number of available interaction periods as 8.

[0107] The number of "available interaction time slots" is a core indicator for measuring the abundance of time resources available for in-depth, non-interference testing of a system. A larger number means more opportunities to conduct sufficient virtual testing to expose problems, thus allowing for more focused analysis of real data. The effect is to transform the abstract concept of resource abundance into a concrete and quantifiable basis for decision-making.

[0108] S42 determines the analysis and processing strategy for the interaction data between the interactive user and the intelligent agent based on the number of available interaction time periods.

[0109] It should be noted that, based on the number of available interaction time slots, the analysis and processing strategy for the interaction data between the user and the agent is determined, and falls into the following categories:

[0110] Case 1: If the number of available interaction time periods is greater than the preset threshold for the number of available time periods, then the analysis and processing strategy for the interaction data between the interactive user and the intelligent agent is determined to be the preset analysis strategy. Only the interaction data of the interactive user who has switched to the human system is analyzed to determine whether there is a deviation in the semantic understanding of the intelligent agent.

[0111] Case 2: If the number of available interaction time periods is not greater than the preset threshold for the number of available time periods, then the analysis and processing strategy for the interaction data between the user and the agent is determined to be the second analysis strategy. In addition to analyzing the interaction data of the user who switched to the human system, it is also necessary to analyze the interaction data with the agent whose number of interactions is greater than the preset threshold for the number of interactions to determine whether there is a deviation in the semantic understanding of the agent.

[0112] Preset available time period threshold: A critical value set by the platform based on experience and resources, such as 10. Preset analysis strategy: A focused and efficient analysis strategy that only analyzes the data that is most likely to have problems.

[0113] The second analysis strategy is an expanded and more comprehensive approach. In addition to core issue data, it analyzes more potentially risky interaction data, including data from interactions that switch to human intervention: these refer to dialogue logs ending with "transferred to human customer service." This is direct evidence of agent processing failure and a high-value sample for analyzing biases.

[0114] Interaction data where the number of interactions with the agent exceeds a preset threshold: This refers to logs showing that the same user (or session) engaged in multiple rounds of dialogue with the agent (e.g., more than 3 interactions) before finally resolving the issue or giving up. This usually indicates inaccurate agent understanding or inefficient guidance, and is an important signal of potential bias.

[0115] Judgment: The number of available time periods is 8 < the preset threshold of 10 → the "greater than" condition is not met. Decision: It belongs to "Case 2" and the "second analysis strategy" is adopted.

[0116] This example demonstrates how to intelligently determine the analysis and processing strategy for real user interaction data based on the system test resource sufficiency index of "available interaction time period," thereby achieving dynamic optimal allocation of operation and maintenance resources.

[0117] Dynamic strategy: The analysis strategy is not fixed, but dynamically adjusted according to the amount of time resources available for proactive testing, reflecting a flexible operation and maintenance philosophy.

[0118] Resource complementarity: A deep understanding that "virtual testing" and "real-world data analysis" are two complementary methods for discovering problems in intelligent agents. When the resources of one method are limited, the strength of the other method is automatically enhanced, ensuring the continuity of problem detection capabilities.

[0119] Focusing on Problem Data: Both strategies avoid the resource black hole of "full log analysis" and always focus on the two types of high-value data that are most revealing of the problem: "transferring to manual processing" and "multiple rounds of inefficiency".

[0120] S3 determines the interaction management strategy for the virtual object in time periods outside the available interaction period based on the analysis and processing results of the interaction data of the virtual object and the interaction data between the interactive user and the intelligent agent during the available interaction period.

[0121] Specifically, the method for determining the interaction management strategy of the virtual object during periods outside the available interaction time is as follows:

[0122] 1. Business Background

[0123] The intelligent customer service system of an e-commerce platform has completed the following tasks: Filtering out "available interaction periods": Through historical load analysis, eight safe periods suitable for large-scale virtual testing (such as the early morning period) were identified, and virtual tests were performed within the "available interaction periods": During these periods, large-scale virtual user dialogue tests were conducted according to the principle of "occupying idle ports at a preset ratio when available".

[0124] Analysis of real interaction data: Based on the number of "available interaction periods" (8, less than the threshold of 10), a second analysis strategy was adopted to deeply analyze the real logs of transitions to human agents and multi-turn dialogues. Current requirement: It is now necessary to develop a strategy to determine whether and how to allow virtual users to interact during other time periods outside of the "available interaction periods" (i.e., when the system is relatively busy or the load is unstable), in order to achieve continuous and gradual monitoring and testing of the intelligent agent.

[0125] 2. Definitions and Status of Core Terms

[0126] Intelligent agent: Intelligent customer service; simulated user: simulated user used for testing; interaction period: 8 selected security testing periods; interacting user: real consumer.

[0127] Analysis and processing strategy: Based on the previous embodiment, the current "second analysis strategy" is adopted (because the number of available time periods 8 < threshold 10). Interaction management strategy: determines the rules for scheduling virtual users during non-safe periods ("outside of available interaction periods"). Preset interaction management strategy: a relatively lenient virtual user scheduling strategy with low trigger conditions. Second preset interaction management strategy: a very strict virtual user scheduling strategy with harsh trigger conditions.

[0128] Preset ratio: The percentage of ports used during the "available interaction period", for example, 20% (i.e., number of virtual users = number of idle ports × 20%). Second preset ratio: A lower percentage of ports used during non-secure periods if a certain management strategy is adopted, for example, 5%. Preset idle number threshold: The minimum number of idle ports required to trigger the "preset interaction management strategy" during non-secure periods, for example, 50. Second preset idle number threshold: The minimum number of idle ports required to trigger a stricter "second preset interaction management strategy" during non-secure periods, for example, 200.

[0129] S51 uses the interaction data of the virtual object during the available interaction period to determine the number of interactions between the virtual object and the agent on different dates;

[0130] The analysis period is the past 7 days (i.e. last week). The above steps have been performed in these 7 days. The number of interactions between virtual objects and intelligent agents on different dates (referring to the number of virtual test dialogues completed each day during the "available interaction period").

[0131] Table 1 Interaction Data

[0132]

[0133] S52 determines the number of interactions where there is a semantic understanding deviation in the interaction data between the interactive user and the intelligent agent based on the analysis and processing results of the interaction data between the interactive user and the intelligent agent, and uses it as the number of understanding deviations.

[0134] In the above steps, the number of semantic misunderstandings refers to the number of interaction cases confirmed to have semantic misunderstandings after analyzing real data using the "second analysis strategy." This data comes from the analysis of real user logs, not virtual tests. Over the past 7 days, the analysis identified 85 real dialogue cases with semantic misunderstandings, which were distributed across various issues, such as "misunderstanding of order-combination rules" and "incorrect return address identification."

[0135] S53 determines the interaction management strategy for the virtual object in time periods outside the available interaction period based on the number of interactions between the virtual object and the agent on different dates, the number of comprehension deviations, and the analysis and processing strategy.

[0136] It is understood that, based on the number of interactions between the virtual object and the agent on different dates, the number of comprehension deviations, and the aforementioned analysis and processing strategy, the interaction management strategy for the virtual object in time periods outside the available interaction period is determined, specifically including:

[0137] S531 Based on the number of interactions between the virtual object and the intelligent agent on different dates, determine whether the number of interactions on different dates is greater than the preset interaction threshold. If so, determine that the interaction management strategy of the virtual object in the time period outside the available time period is no interaction processing required. If not, proceed to the next step.

[0138] In the above steps, check the sufficiency and stability of the virtual test. Judgment: Check whether the number of virtual interactions per day for the past 7 days is greater than 1000 times. Result: Friday's data is 800 times, which is not greater than 1000 times. Decision: The condition of "greater than" is not met → Proceed to S532.

[0139] The requirement for virtual testing to reach a certain scale daily is to ensure the continuity and consistency of testing pressure. Failure to meet this target on Friday indicates fluctuations in testing, meaning existing test results cannot be relied upon entirely, and therefore, the decision to "avoid interaction during unsafe periods" cannot be made directly.

[0140] S532 takes the date with no more than a preset interaction number threshold as the interaction deviation date, and determines whether the proportion of the interaction deviation date in the date after the use of the available interaction period is greater than the preset interaction deviation date number threshold. If so, the interaction management strategy of the virtual object in the period outside the available period is determined to be the preset interaction management strategy. If not, proceed to the next step.

[0141] In the above steps, assess the percentage of days with insufficient testing and identify interaction deviation days: only Friday (800 times) is a day with no more than the threshold number of interactions.

[0142] Percentage calculated: Number of interaction deviation dates = 1 day. These dates are after the "adoption of available interaction period" (i.e., last week). Total number of dates = 7 days, percentage = 1 / 7 ≈ 14.3%.

[0143] Judgment: 14.3% is not greater than the preset interaction deviation date quantity threshold (50%), and does not meet the "greater than" condition → proceed to S533.

[0144] While there were days when testing was insufficient, this was not a common occurrence (only 14.3%). Therefore, instead of taking immediate action, it is necessary to consider more direct agent performance metrics (i.e., understanding bias rate) for judgment.

[0145] S533 uses the proportion of the number of comprehension deviations in the number of interactions in the analysis and processing of interactive data as the interaction deviation proportion, and determines whether the interaction deviation proportion is greater than a preset interaction deviation proportion threshold. If so, the interaction management strategy of the virtual object in the time period outside the available time period is determined to be the preset interaction management strategy. If not, proceed to the next step.

[0146] In the above steps, the severity of the agent's current comprehension bias is checked, and the percentage of interaction bias is calculated: as calculated before, it is approximately 5.67%.

[0147] Judgment: 5.67% is not greater than the preset interaction deviation percentage threshold (8%). Decision: The "greater than" condition is not met → proceed to S534. The semantic understanding error rate (5.67%) exhibited by the agent in actual service is within a controllable range and has not reached the severity level (8%) that requires high vigilance. Therefore, the strict preset interaction management strategy is not activated.

[0148] S534 determines whether the percentage of the interaction deviation is less than a preset deviation percentage threshold. If yes, it determines that the interaction management strategy of the virtual object in the time period outside the available time period is no interaction processing required. If no, it proceeds to the next step.

[0149] In the above steps, the performance of the agent is checked to determine if the interaction deviation ratio (5.67%) is not less than the preset deviation ratio threshold (2%). Decision: The "less than" condition is not met → proceed to S535. Although the deviation rate is not serious, it is not low enough to be negligible (2%). Therefore, monitoring during non-safe periods cannot be completely abandoned, and the final decision needs to be made based on the current depth of analysis.

[0150] S535 acquires the analysis and processing strategy, determines whether the analysis and processing strategy is a preset analysis strategy, if yes, then determines the interaction management strategy of the virtual object in the time period outside the available time period as a preset interaction management strategy, if no, then determines the interaction management strategy of the virtual object in the time period outside the available time period as a second preset interaction management strategy.

[0151] In the above steps, a final decision is made based on the depth of the analysis and processing strategy to obtain the current analysis and processing strategy: as mentioned above, it is the "second analysis strategy". It is determined that this strategy is not the "preset analysis strategy". The final decision (according to the S535 logic) is to determine the interaction management strategy of the virtual object in the time period outside the available time period as the second preset interaction management strategy.

[0152] The current approach employs a "secondary analysis strategy," meaning we are already deeply mining real user data to uncover issues. Since real data monitoring is already quite thorough, to avoid overloading the system with excessive testing, a more stringent and conservative scheduling strategy (i.e., the secondary preset interaction management strategy) should be adopted for virtual users during non-safety periods. Conversely, if a "preset analysis strategy" (analyzing only manually converted data) is used, it indicates a coarser level of problem monitoring. In this case, virtual users would need to perform relatively more supplementary probing during non-safety periods (i.e., the preset interaction management strategy).

[0153] Furthermore, the preset interaction management strategy is to determine the number of interactions between the virtual user and the intelligent agent when the number of idle ports during the time period is greater than a preset idle number threshold.

[0154] The preset interactive management strategy has the following triggering and execution rules: Condition A (relatively abundant resources): number of idle ports > preset idle number threshold (50 ports); Condition B (low occupancy rate): number of virtual users = number of idle ports × second preset ratio (5%, note that the second preset ratio is the same as the previous strategy, but the trigger threshold is lower).

[0155] Furthermore, the second preset interaction management strategy is to determine the number of interactions between the virtual user and the intelligent agent by using a second preset proportion of the number of idle ports when the number of idle ports during the time period is greater than a second preset idle number threshold.

[0156] The second preset interaction management strategy stipulates that, in any non-interaction-available period (e.g., peak daytime hours on weekdays), virtual user interactions will only be triggered if both of the following conditions are met simultaneously:

[0157] Condition A (Abundant Resources): During the current period, the number of idle ports of the intelligent agent system is greater than the second preset idle number threshold (200). Condition B (Extremely Low Occupancy Rate): Once triggered, the number of virtual users allowed to start = number of idle ports × second preset rate (5%).

[0158] Furthermore, the second preset ratio is less than the preset ratio, and the second preset idle quantity threshold is greater than the preset idle quantity threshold.

[0159] Furthermore, during the available interaction period, as long as there are idle ports, the number of interactions between virtual users and intelligent agents is determined according to a preset proportion of idle ports.

[0160] Example 2

[0161] In a second aspect, the present invention provides a computer system comprising: a memory and a processor connected in communication, and a computer program stored in the memory and capable of running on the processor, wherein the processor executes the above-described method for training and managing an intelligent agent when running the computer program.

[0162] Example 3

[0163] Thirdly, this application provides a training management device for an intelligent agent, employing the aforementioned training management method for an intelligent agent, specifically including:

[0164] The module includes an interaction time segmentation module, an analysis strategy determination module, and an interaction management module.

[0165] The interaction time segmentation module is responsible for determining whether the interaction request type between the virtual user and the intelligent agent belongs to the preset request type.

[0166] The analysis strategy determination module is responsible for determining the analysis and processing strategy for the interaction data between the interactive user and the intelligent agent.

[0167] The interaction management module is responsible for determining the interaction management strategy for virtual objects during periods outside of the available interaction time slots.

[0168] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the embodiments of apparatus, devices, and non-volatile computer storage media are basically similar to the method embodiments, so the descriptions are relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0169] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0170] The above description is merely one or more embodiments of this specification and is not intended to limit this specification. Various modifications and variations can be made to the one or more embodiments of this specification by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of one or more embodiments of this specification should be included within the scope of the claims of this specification.

Claims

1. A training and management method for intelligent agents, characterized in that, Specifically, it includes: The virtual user is constructed based on the interaction data of the intelligent agent. Based on the interaction data between the intelligent agent and the interactive user, the matching situation of the interaction needs between the intelligent agent and the interactive user is determined. Based on the matching situation and the interactive user data in different time periods, if it is determined that the interaction need type between the virtual user and the intelligent agent does not belong to the preset need type, the next step is performed. Based on the interaction needs type between the virtual user and the intelligent agent, and combined with the changes in the interaction user data in different time periods, the available interaction time periods between the virtual user and the intelligent agent are determined. Based on the available interaction time period data, the analysis and processing strategy for the interaction data between the interaction user and the intelligent agent is determined. Based on the analysis and processing results of the interaction data of virtual users during the available interaction period and the interaction data between interactive users and intelligent agents, the interaction management strategy of the virtual users in the period outside the available interaction period is determined. Determining that the interaction request type between the virtual user and the intelligent agent does not belong to the preset request type specifically includes: Based on the matching results, the number of times the agent and the user are switched to the human system during the interaction process is determined and used as the switching count; Based on the interactive user data of the agent in different time periods, the total duration of idle ports of the agent in different time periods is determined and used as the idle duration; Based on the number of switching times and the idle time in different time periods, it is determined whether the interaction demand type between the virtual user and the intelligent agent belongs to the preset demand type.

2. The training and management method for intelligent agents as described in claim 1, characterized in that, The virtual user is a virtual user constructed based on the interaction data between the intelligent agent and the interactive user. The virtual user and the intelligent agent interact with each other to identify the intelligent agent's recognition deviation of the interaction data.

3. The training and management method for intelligent agents as described in claim 1, characterized in that, The interactive user is the user who interacts with the intelligent agent to exchange data.

4. The training and management method for intelligent agents as described in claim 1, characterized in that, The matching of the interaction needs between the intelligent agent and the interactive user is determined based on the number of times the interaction between the intelligent agent and the interactive user is transferred to the human system during the interaction process.

5. The training and management method for intelligent agents as described in claim 1, characterized in that, If the interaction requirement between the virtual user and the intelligent agent is a preset requirement type, then the number of interactions between the virtual user and the intelligent agent will be determined according to a preset proportion of the number of idle ports, provided that there are idle ports in different time periods.

6. The training and management method for intelligent agents as described in claim 5, characterized in that, The preset ratio is between 0 and 1.

7. The training and management method for intelligent agents as described in claim 1, characterized in that, The method for determining the interaction management strategy for virtual users outside of the available interaction time period is as follows: Based on the interaction data of virtual users during the available interaction period, determine the number of interactions between virtual users and the intelligent agent on different dates; Based on the analysis and processing results of the interaction data between the interactive user and the intelligent agent, the number of interactions in which the semantic understanding of the interaction data between the interactive user and the intelligent agent is determined, and this number is taken as the number of understanding deviations. Based on the number of interactions between the virtual user and the agent on different dates, the number of comprehension deviations, and the analysis and processing strategy, the interaction management strategy for the virtual user in time periods outside the available interaction period is determined.

8. A computer system, comprising: A memory and a processor connected in communication, and a computer program stored in the memory and capable of running on the processor, characterized in that, when the processor runs the computer program, it executes a training management method for an intelligent agent as described in any one of claims 1-7.

9. A training and management device for an intelligent agent, employing the training and management method for an intelligent agent as described in any one of claims 1-7, characterized in that, Specifically, it includes: The module includes an interaction time segmentation module, an analysis strategy determination module, and an interaction management module. The interaction time segmentation module is responsible for determining whether the interaction request type between the virtual user and the intelligent agent belongs to the preset request type. The analysis strategy determination module is responsible for determining the analysis and processing strategy for the interaction data between the interactive user and the intelligent agent. The interaction management module is responsible for determining the interaction management strategy for virtual users during periods outside of the available interaction time slots.

Images