Multi-agent cognitive strategy training and evaluation method, system and device and medium

Abstract

The invention relates to the technical field of artificial intelligence, can be applied to medical auxiliary diagnosis, financial risk analysis and other scenes, and provides a multi-agent cognitive strategy training and evaluation method, system and device and a medium, and the method comprises the steps: constructing a teacher agent and a plurality of learner agents with different learning behavior characteristics; running a simulation teaching cycle, controlling the teacher agent to publish learning stage information based on a preset scene problem, and controlling each learner agent to execute the information to generate an interaction response containing process behavior data and learning result data; recording process behavior data and learning result data of each agent, and generating a cognitive strategy evaluation report of each agent based on the process behavior data and the learning result data; and combining the plurality of learner agents into a multi-agent cluster capable of cooperatively solving the preset scene problem according to strategy features indicated by the evaluation report. According to the invention, dynamic simulation and evaluation of the agent cognitive strategy are realized, and a cooperative solution cluster oriented to an actual scene can be constructed according to the dynamic simulation and evaluation.

Description

Technical Field

[0001] This disclosure relates to the field of artificial intelligence technology, and more specifically, to a method, system, device, and medium for training and evaluating multi-agent cognitive strategies. Background Technology

[0002] With breakthroughs in natural language understanding and task generation achieved by large-scale artificial intelligence models, educational technology is moving from "content delivery" to a deeper intelligent stage of "cognitive modeling" and "strategic intervention." This means that systems can not only transmit knowledge but also simulate and respond to learners' thought processes and developmental states, thereby achieving highly personalized and adaptive business support. For example, in complex decision-making fields such as medical assisted diagnosis and financial risk analysis, there is an urgent need to obtain intelligent agents with human expert cognitive strategies through simulation training, and to evaluate their decision robustness and strategy effectiveness in different contexts.

[0003] However, current methods in the field of agent cognitive simulation and training mainly rely on static task testing, single evaluation metrics, and fixed behavioral rules. While these methods can train agents to complete specific tasks, they generally suffer from limitations such as rigid training processes, untraceable policy evolution paths, and single evaluation dimensions. These methods typically treat agents as execution entities with fixed behavioral patterns, making it difficult to characterize the dynamic adaptation and evolution of their cognitive strategies in long-term, multi-stage tasks. Furthermore, related training systems generally lack a systematic construction and quantitative evaluation of agents' differentiated cognitive characteristics (such as decision-making depth, reflective tendencies, and collaborative styles), resulting in trained agents that struggle to perform collaboratively in real-world, complex scenarios. Summary of the Invention

[0004] This disclosure provides at least one method, system, device, and medium for training and evaluating multi-agent cognitive strategies. By constructing a multi-agent collaborative architecture, it simulates the interaction process between teachers and different types of learners in real teaching scenarios. This enables more accurate capture of the dynamic behavioral characteristics and cognitive change patterns of each learner during the learning process, achieving dynamic simulation and refined evaluation of agent cognitive strategies. Based on this, it can construct a collaborative solution cluster oriented towards real-world scenarios.

[0005] This disclosure provides a method for training and evaluating multi-agent cognitive strategies, including: Construct and initialize a teacher agent and multiple learner agents; wherein the learner agents are configured as types with different learning behavior characteristics, including deep learners, surface learners, lazy learners and general learners; A simulated teaching cycle is run, in which the teacher agent is controlled to publish learning stage information to all learner agents; and the learner agents are controlled to execute the learning stage information based on their respective configured types, and generate interactive response information in response to the learning stage information; wherein the learning stage information is determined based on a preset scenario problem, and the interactive response information includes process behavior data and learning result data; Record the process behavior data and learning outcome data of each learner agent during the simulated teaching cycle; and generate a cognitive strategy diagnostic report for each learner agent based on the process behavior data and the outcome data. Based on the policy characteristics indicated by the cognitive policy evaluation reports of each learner agent, the multiple learner agents are combined to form a multi-agent cluster capable of collaboratively solving the preset scenario problem.

[0006] This disclosure provides a multi-agent cognitive strategy training and evaluation system, including: The agent building unit is used to build and initialize a teacher agent and multiple learner agents with different learning behavior characteristics, including deep learners, surface learners, lazy learners and general learners; The cycle simulation unit is used to run a simulated teaching cycle, control the teacher agent to release learning stage information, and control the learner agent to execute the learning stage information and generate interactive responses to the learning stage information; wherein, the learning stage information is determined based on a preset scenario problem, and the interactive response information includes process behavior data and learning result data; The diagnostic report generation unit is used to record the process behavior data and learning outcome data of each learner agent in the simulated teaching cycle; and to generate a cognitive strategy diagnostic report for each learner agent based on the process behavior data and the outcome data. The agent collaboration construction unit is used to combine the multiple learner agents based on the policy features indicated by the cognitive policy evaluation reports of each learner agent, so as to form a multi-agent cluster capable of collaboratively solving the preset scenario problem.

[0007] This disclosure provides a multi-agent cognitive strategy training and evaluation device, including: The agent construction module is used to construct and initialize a teacher agent and multiple learner agents; wherein, the learner agents are configured as types with different learning behavior characteristics, including deep learners, surface learners, lazy learners and general learners; The teaching simulation module is used to run a simulated teaching cycle. During the simulated teaching cycle, the teacher agent is controlled to publish learning stage information to all learner agents. The learner agents are also controlled to execute the learning stage information based on their respective configured types and generate interactive response information in response to the learning stage information. The learning stage information is determined based on a preset scenario problem, and the interactive response information includes process behavior data and learning result data. The report generation module is used to record the process behavior data and learning outcome data of each learner agent in the simulated teaching cycle; and to generate a cognitive strategy diagnostic report for each learner agent based on the process behavior data and the outcome data. The agent combination module is used to combine the multiple learner agents based on the policy characteristics indicated by the cognitive policy evaluation reports of each learner agent, so as to form a multi-agent cluster that can collaboratively solve the preset scenario problem.

[0008] This disclosure provides a computer device, including a processor, a memory, and a bus. The memory stores machine-readable instructions executable by the processor. When the computer device is running, the processor communicates with the memory via the bus. When the machine-readable instructions are executed by the processor, they perform the multi-agent cognitive strategy training and evaluation method as described in any of the above possible embodiments.

[0009] This disclosure provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the multi-agent cognitive strategy training and evaluation method as described in any of the possible embodiments above.

[0010] The multi-agent cognitive strategy training and evaluation method, system, device, and medium provided in this disclosure, by constructing a teacher agent and multiple learner agents with different learning behavior characteristics, and running a simulated teaching cycle to control their release and execution of learning stage information and generation of interactive response information, achieves parallel simulation and data collection of the behaviors and results of different types of learners. By recording the process behavior data and learning result data generated by each learner in the cycle, and generating a cognitive strategy diagnostic report based on this, it achieves synchronous tracking and quantitative analysis of the cognitive learning processes of multiple differentiated individuals, thereby revealing the characteristics and effects of different cognitive strategies from group behavior comparison and individual data accumulation. Furthermore, by screening and combining agents according to the strategy characteristics indicated by the evaluation report, it can transform the deep understanding of agent strategies obtained in the training and evaluation process into a basis for constructing a multi-agent cluster with clear role division and collaborative mechanisms for specific scenarios, realizing a closed loop from individual ability assessment to group effectiveness optimization.

[0011] Thus, by constructing a multi-agent collaborative architecture, this embodiment of the present disclosure can simulate the dynamic evolution process of the cognitive strategies of agents in long-term, multi-stage tasks, and achieve a refined evaluation of the cognitive strategies of agents; and based on the policy characteristics obtained from the evaluation, agents with complementary strategies are combined into a collaborative cluster, thereby constructing a multi-agent cluster that can efficiently and robustly solve problems in various scenarios.

[0012] To make the above-mentioned objects, features and advantages of this disclosure more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0013] To more clearly illustrate the technical solutions of the embodiments of this disclosure, the accompanying drawings referenced in the embodiments will be briefly described below. These drawings are incorporated in and constitute a part of this specification. They illustrate embodiments conforming to this disclosure and, together with the specification, serve to explain the technical solutions of this disclosure. It should be understood that the following drawings only show some embodiments of this disclosure and should not be considered as limiting the scope. Those skilled in the art can obtain other related drawings based on these drawings without creative effort.

[0014] Figure 1 This diagram illustrates an application environment of a multi-agent cognitive strategy training and evaluation method provided in an embodiment of this disclosure. Figure 2 A flowchart of a multi-agent cognitive strategy training and evaluation method provided in an embodiment of this disclosure is shown; Figure 3A flowchart of a simulated teaching cycle operation method provided by an embodiment of this disclosure is shown; Figure 4 A flowchart illustrating a method for constructing a long-term memory bank for a learner agent provided in an embodiment of this disclosure is shown. Figure 5 A flowchart of a self-assessment method provided by an embodiment of this disclosure is shown; Figure 6 A flowchart of a cognitive depth assessment method provided by an embodiment of this disclosure is shown; Figure 7 This diagram illustrates the structure of a multi-agent cognitive strategy training and evaluation system provided in an embodiment of this disclosure. Figure 8 This diagram illustrates the structure of a multi-agent cognitive strategy training and evaluation device provided in an embodiment of the present disclosure. Figure 9 A schematic diagram of the structure of a computer device provided in an embodiment of this disclosure is shown. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of the embodiments of this disclosure clearer, the technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this disclosure, and not all of them. The components of the embodiments of this disclosure described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this disclosure provided in the accompanying drawings is not intended to limit the scope of the claimed disclosure, but merely represents selected embodiments of this disclosure. All other embodiments obtained by those skilled in the art based on the embodiments of this disclosure without inventive effort are within the scope of protection of this disclosure.

[0016] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0017] In this document, the term "and / or" merely describes a relationship, indicating that three relationships can exist. For example, A and / or B can represent three cases: A alone, A and B simultaneously, and B alone. Furthermore, the term "at least one" in this document means any combination of at least two of any one or more elements. For example, including at least one of A, B, and C can mean including any one or more elements selected from the set consisting of A, B, and C.

[0018] To facilitate understanding of this embodiment, the executing entity of the multi-agent cognitive strategy training and evaluation method provided in this disclosure embodiment will first be described in detail. The multi-agent cognitive strategy training and evaluation method provided in this invention embodiment can be applied to, for example... Figure 1 In this application environment, the client communicates with the server via a network. The client can be a mobile device, user terminal, terminal, handheld device, computing device, etc. The server can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud storage, big data, and artificial intelligence platforms.

[0019] The multi-agent cognitive strategy training and evaluation method provided in this application embodiment will be described in detail below with reference to the accompanying drawings. See also Figure 2 The diagram shows a flowchart of a multi-agent cognitive strategy training and evaluation method provided in this embodiment of the present disclosure. The method includes the following steps S201-S204: S201, construct and initialize a teacher agent and multiple learner agents.

[0020] As is understood, an intelligent agent refers to an autonomous computational entity implemented by a computer program, capable of simulating the behavior of a specific role or entity, and interacting with other intelligent agents or the environment to complete tasks. It can be constructed based on a pre-trained large-scale language model and endowed with specific behavioral logic and knowledge background through prompting engineering or fine-tuning techniques. In this disclosure, the teacher intelligent agent is an intelligent programmable entity that simulates the role of a teacher in a real teaching scenario. It possesses capabilities such as issuing learning tasks, guiding the learning process, and evaluating learning outcomes. For example, in the medical field, the teacher intelligent agent can be set as an experienced medical expert intelligent agent, capable of formulating reasonable learning plans for the learner intelligent agent based on the medical knowledge system and learning objectives.

[0021] Here, a learner agent refers to a type that simulates a real learner and is configured with different learning behavioral characteristics. These characteristics can include learning speed, learning style, knowledge mastery, cognitive strategies, and other aspects. For example, some learner agents can be set to learn quickly but have a less solid grasp of knowledge. When faced with new knowledge, they can absorb it rapidly, but they are prone to gaps in their understanding and application. Other learner agents can be set to learn slowly but have a rigorous and meticulous learning style, striving to understand each knowledge point thoroughly, and thus have a solid grasp of knowledge. Furthermore, a learner agent can be configured to have uneven knowledge mastery, with strong knowledge in some technical areas but relatively weak knowledge in others, and their learning style may be proactive at times and passive and slack when encountering difficulties at others.

[0022] In some possible implementations, learner agents can be categorized into deep learners, surface learners, lazy learners, and general learners based on different learning behavior characteristics, reflecting individuals with different levels of motivation, self-awareness, and cognitive strategies. Deep learners typically exhibit strong intrinsic motivation, tending to delve into the underlying principles and logical connections of knowledge and perform cross-contextual knowledge transfer; surface learners rely more on memory and pattern recognition, tending to master standard processes and surface features, lacking an understanding of deep structures; lazy learners exhibit low learning engagement, often choosing the easiest path and requiring external incentives or structured guidance; general learners serve as a baseline model, without pre-defined behavioral tendencies, used to observe and record default learning patterns without specific guidance. For example, in the financial field, deep learners may correspond to analysts who focus on in-depth fundamental research and explore macroeconomic settings with different risk preferences and financial knowledge bases. Surface learners may correspond to traders who mainly rely on technical indicators, charts, and historical patterns for trading. Lazy learners may correspond to investors who tend to follow market trends or others' advice and lack independent research. General learners can be used to simulate the natural learning trajectory of a novice with no professional background in the market in an initial state.

[0023] In some possible embodiments, the learner agent may also set different learning behavior feature dimensions and parameter combinations according to the teaching objectives and assessment needs of the actual simulation scenario, without making specific limitations here.

[0024] S202, run a simulated teaching cycle, in which the teacher agent is controlled to publish learning stage information to all learner agents; and the learner agents are controlled to execute the learning stage information based on their respective configured types, and generate interactive response information for the learning stage information.

[0025] Specifically, a simulated teaching cycle refers to a complete, structured simulated teaching process over a period of time, which may include multiple stages arranged in a logical order. Learning stage information refers to the specific instructions or tasks issued by the teacher agent within this cycle. This information can be determined based on the pre-set scenario problem and may include learning content, learning objectives, and learning time arrangements. The pre-set scenario refers to a problem situation in a specific domain designed to be solved through training and evaluating the agent. This could be a high-frequency trading market volatility risk identification and response strategy deduction in the financial field, or a comprehensive diagnosis of rare diseases based on multimodal medical images and patient history in the medical field; it could also be complex production line fault diagnosis and collaborative scheduling decisions in the industrial field, dynamic path planning for multi-vehicle collaboration in extreme weather in the autonomous driving field, or multi-step attack chain identification and collaborative defense strategy generation in the cybersecurity field, among other task scenarios, without specific limitations.

[0026] Here, learning content refers to the specific knowledge units or problem situations that the teacher agent requires the learner agent to master or analyze. For example, in a medical simulation teaching scenario, it might be a complex medical record containing detailed information such as patient symptoms, examination results, and medical history. Learning objectives refer to the ability standards or outcome indicators that the teacher agent expects learners to achieve after completing this stage. For example, in financial investment simulation teaching, students are required to master the application of specific investment strategies and achieve a certain accuracy rate in predicting returns. Learning time arrangement refers to the teacher agent's time-series planning for each stage of the learning process, clarifying the time allocation for each learning task, such as allocating time for analyzing specific investment decision problems, to ensure the orderly progress of teaching.

[0027] Furthermore, the learner's intelligent experience, based on its configured type, executes learning phase information and generates interactive response information. The interactive response information is the learner's output after processing this information. It includes process behavior data reflecting their thinking and decision-making processes, such as which examination indicators were reviewed when diagnosing a case, or whether sensitivity analysis was conducted during the decision-making process. Learning outcome data refers to the quantifiable output after the task is completed, such as the final diagnostic conclusion or the predicted rate of return on investment recommendations.

[0028] Here, different learner agents, due to their varying cognitive strategies and motivational levels, will correspond to different execution paths and data output patterns. Taking medical simulation teaching as an example, when the teacher agent presents a case of pneumonia with rare complications, the deep learner agent, during execution, will utilize its powerful analytical and reasoning abilities. The corresponding process behavior data will meticulously record the differentiation and deduction of various similar symptoms. By consulting medical literature and referring to past cases, it will organize its diagnostic thinking and ultimately output a personalized treatment plan, forming an interactive response information containing both process behavior data and learning outcome data. The basic learner agent, on the other hand, focuses on the application of basic knowledge, analyzing cases according to basic diagnostic procedures. Its process behavior data mainly revolves around the interpretation of basic examination indicators and the application of basic diagnostic methods, outputting relatively conventional treatment plans.

[0029] For example, the specific operation of a simulated teaching cycle can follow a fixed sequence of stages to support the accumulation of long-term, comparable behavioral data, referring to... Figure 3 As shown, running a simulated teaching cycle may include the following steps S301~S305: S301, the simulated teaching cycle is divided into multiple consecutive simulated time periods.

[0030] Each simulation time period represents a complete learning cycle and contains multiple consecutive simulation time units executed sequentially. For example, the entire simulation can last for twelve simulation time periods, with each simulation time period containing four simulation time units, corresponding to the four stages of knowledge input, task practice, test evaluation, and summary debate.

[0031] S302, in the first simulation time unit, the teacher agent is controlled to publish knowledge teaching content, and each learner agent is controlled to receive the knowledge teaching content.

[0032] Here, the knowledge instruction content refers to the collection of core concepts, principles, methods, or factual information systematically organized by the teacher agent according to the teaching syllabus and objectives, intended to be transmitted to the learner agent. The first simulation time unit can be designed as the knowledge instruction phase of the entire simulation cycle. In this phase, the teacher agent imparts one or more key knowledge points to the learner agent through structured presentations, case analyses, or multimedia material presentations. For example, in a financial investment simulation, the teacher agent might teach the theoretical derivation and assumptions of the Capital Asset Pricing Model; in a medical diagnosis simulation, it might systematically explain the typical pathophysiological mechanisms and clinical manifestations of a certain disease.

[0033] Furthermore, the learner agent is controlled to receive this information and understand and internalize it based on its configured cognitive type and existing knowledge background. Here, during the reception process, the deep learner agent tends to actively connect with existing knowledge, question hypotheses, and perform deep encoding; its reception behavior data may include question records and knowledge association notes. In contrast, the surface learner agent focuses more on memorizing key conclusions and formulas; its behavior data mainly consists of highlighting and retelling key points.

[0034] S303, in the second simulation time unit, the teacher agent is controlled to issue the learning task corresponding to the knowledge teaching content, and each learner agent is controlled to execute the learning task.

[0035] Here, learning tasks refer to specific problems, exercises, or practical projects designed by the teacher agent. Their purpose is to encourage learner agents to apply the received knowledge, consolidate their understanding, and cultivate application skills through practical operation. In the second simulation time unit, the learning tasks issued by the teacher agent are usually directly related to the knowledge taught in the previous unit, forming a "learn-practice" integrated teaching link. The task formats can be diverse. For example, in a financial scenario, learner agents might be required to use the taught pricing model to value a group of real listed companies; in a medical scenario, they might be required to perform a preliminary diagnostic analysis of a simplified medical record based on the taught pathological mechanisms. Each learner agent performs the task according to its type: deep learners will try to verify the model from multiple angles, consider boundary conditions, or explore differential diagnoses; surface learners may strictly follow the taught steps, pursuing correct output format; lazy learners may only complete the minimum requirements or seek simplified paths. During this process, the problem-solving steps, knowledge units invoked, methods tried, and intermediate results of each learner agent are recorded to constitute the process behavior data of each learner agent.

[0036] S304, in the third simulation time unit, the teacher agent is controlled to issue test tasks and trap questions, receive the answers of each learner agent to the test tasks and trap questions, and issue feedback information on the correctness of the answers to each learner agent based on the preset answers.

[0037] Specifically, test tasks are standardized questions used to routinely assess learners' mastery of the taught content, and their answers can usually be derived directly from the learned knowledge. Trap questions, on the other hand, are specially designed diagnostic questions that are highly similar to test tasks in their surface structure, expression, or problem context, but whose core logic, key premises, or implicit conditions have undergone meaningful and subtle changes. Directly applying knowledge or models will not yield correct answers. For example, in financial valuation tests, the test task might apply a model under standard market conditions, while trap questions might apply the model to a market environment with liquidity shortages or special regulatory policies. In medical diagnostic tests, the test task might identify typical symptoms, while trap questions might combine typical symptoms with drug side effects or rare comorbidities.

[0038] Furthermore, based on the preset correct answers, instant feedback information can be sent to each learner agent, clearly indicating whether their answer is correct. Here, the feedback information is not only an important part of the teaching loop, but its content (such as correctness and reaction time) will also be recorded as key learning outcome data for subsequent comprehensive diagnostic analysis.

[0039] In this way, by receiving and comparing the answers of learner agents to these two types of questions, it is possible to effectively identify whether each learner agent truly understands the internal logic of the knowledge and can apply it flexibly, or merely relies on surface features for pattern matching.

[0040] It is understandable that learners' self-regulation ability is also an important part of their cognitive development and strategy evolution in the complete teaching process. Since continuous knowledge input and task execution during the simulated teaching cycle will consume cognitive resources and may cause learning fatigue or motivational fluctuations, in order to more realistically reflect the behavioral logic of each agent and obtain their strategy preference data, a self-regulation link can be introduced after the key evaluation node. Specifically, it can include the following steps (1) to (4): (1) Send a strategy selection instruction to each of the learner agents respectively; wherein the strategy selection instruction provides a summary option and a rest option; (2) Receive the selection results returned by each learner agent based on its configured type; (3) If the result is the summary option, then a learning summary record is generated and stored for the learner agent; (4) If the selected result is the rest option, then record a rest behavior for the learner agent and skip the step of generating a learning summary record.

[0041] Specifically, strategy selection instructions are typically triggered after completing a cognitively challenging phase (such as testing and feedback), simulating learners' autonomous time management after completing a phase of learning. The "summary" option represents the learner's choice to actively review and organize the knowledge, performance, and errors of the recently completed phase; the "rest" option represents the learner's choice to pause learning activities for rest. Learner agents with different behavioral characteristics will make different choices based on their intrinsic motivation and strategic inclinations. Continuing with the example of the learner agent mentioned above, the deep learner agent, due to its high motivation and reflective habits, is more likely to choose the summary option; the lazy learner agent may choose the rest option more frequently due to its low engagement; the surface learner agent may waver in different situations; and the choices of the general learner agent can be used to observe behavioral patterns without pre-set preferences.

[0042] Here, if a learner agent chooses the summary option, a learning summary record can be generated and stored for that learner agent. This record is an active summary of the agent's learning process and results from the previous stage, and its content may include a review of knowledge points, an analysis of the reasons for incorrect answers, and an evaluation of its own learning methods. For example, in a medical simulation, a surgeon agent who chooses the summary option might generate a record analyzing why it overlooked a patient's history of drug allergies in a tricky diagnosis and formulating key points for subsequent consultations.

[0043] Here, if a learner agent chooses the rest option, a rest action can be recorded for that learner agent, skipping the step of generating a learning summary record. At this point, the selection event and its timing will be recorded. Continuous or frequent rest actions themselves become key process behavior data, potentially correlated with subsequent declines in learning effectiveness or fluctuations in motivation. For example, for an analyst agent that continuously chooses rest, the corresponding cognitive strategy diagnostic report might indicate a tendency towards insufficient learning engagement.

[0044] S305, in the fourth simulation time unit, the teacher agent is controlled to organize the learner agents to participate in a debate task on a preset topic, and the speech content and opinion change data of each learner agent in the debate task are recorded.

[0045] Specifically, the fourth simulated time unit refers to the stage within the simulated teaching cycle dedicated to social cognitive interaction and collective knowledge construction. In this unit, the teacher agent no longer unilaterally imparts knowledge or assigns independent tasks, but instead acts as a discussion organizer and coordinator, proposing a pre-set topic that is controversial, open-ended, or has multiple interpretations, guiding all learner agents to engage in structured debate around this topic. The content of the speech refers to all textual or semantic expressions expressed by learners during the debate, including their arguments, cited evidence (such as previously taught concepts and data derived from tasks), logical reasoning, and questions or responses to others. The data on changes in viewpoints refers to the quantitative or qualitative tracking records of the learner agents' core positions at different points in the debate process, used to characterize whether their cognitive positions shift during the interaction, the timing of these shifts, and the possible influencing factors.

[0046] For example, in a financial investment simulation, the pre-set topic could be "Given the current macroeconomic warning signals, should we prioritize increasing holdings in defensive sectors or aggressive growth stocks?" The deep learner agent's statements might include comparative analysis of historical cycle data, assessments of the cash flow resilience of different industries, and, after hearing strong arguments about the possibility of policy stimulus, its viewpoint change data might show a gradual shift from a firm advocacy for defensive strategies to a more flexible stance accepting some growth stock allocations. In contrast, the surface learner agent's statements might more often repeat classic asset allocation principles taught by the instructor, and its viewpoint change data might show a tendency to waver or simply conform to the majority opinion.

[0047] In this way, by recording the logical quality, evidence relevance, and evolution path of these statements, the key variable of social interaction can be incorporated into cognitive modeling. This allows for the assessment of learners' cognitive resilience, critical thinking, and knowledge integration abilities when faced with conflicting information, group pressure, or persuasive communication. Data from these dimensions enriches the in-depth diagnosis of each learner's cognitive strategies in subsequent steps.

[0048] Here, the knowledge instruction content, learning tasks, test tasks, trap questions, and debate tasks on preset topics mentioned above all belong to the learning stage information published by the teacher agent. In some other embodiments, other types of learning activities or assessment tasks may also be included, such as collaborative projects, practical operation simulations, open case analysis, etc., which are not specifically limited here.

[0049] In this embodiment of the disclosure, by constructing a structured, long-term simulated teaching closed loop, dynamic tracking and data collection of the entire process from knowledge imparting, practical application, in-depth assessment to social interaction and self-regulation are realized. Together, they provide a comprehensive, three-dimensional and time-series behavioral data foundation for generating in-depth and personalized cognitive strategy diagnostic reports, and realize dynamic, in-depth and interpretable diagnosis of the cognitive abilities of individuals or models.

[0050] In some other embodiments, the simulated teaching cycle can be divided into basic cycle units based on different time scales such as weeks, bi-monthly or quarterly. Each basic cycle unit can also be designed with different numbers and different combinations of simulated time units, which are not specifically limited here.

[0051] S203, record the process behavior data and learning outcome data of each learner agent in the simulated teaching cycle; and generate a cognitive strategy diagnostic report for each learner agent based on the process behavior data and the outcome data.

[0052] Understandably, process behavior data refers to the collection of behavioral sequences, choice paths, and interaction traces generated by the learner agent during cognitive activities and decision-making operations within a simulated teaching cycle; learning outcome data refers to the quantifiable or evaluable final outputs and performance indicators generated by the learner agent after completing learning tasks, tests, or debates at each stage. Here, to achieve long-term, coherent tracking of learners' cognitive states and support their adaptive learning by recalling past experiences in new tasks, a continuously updated personalized memory bank can be built for each learner during the simulated teaching cycle. (Refer to...) Figure 4 As shown, the following steps S401~S403 may be included when recording process behavior data and learning result data: S401 establishes and maintains a long-term memory bank for each learner agent.

[0053] Here, the long-term memory is a structured storage module uniquely bound to each learner agent, used to persistently store all the key data generated during the simulated learning process.

[0054] S402, for each learner agent, the process behavior data and learning result data of the learner agent, as well as the simulation time unit identifier corresponding to each data, are stored in the long-term memory bank corresponding to the learner agent.

[0055] Specifically, during the learning process, each learner agent generates process behavior data and interim learning result data in real time. This data can be associated with the corresponding simulation time unit identifier and stored in its dedicated long-term memory. The simulation time unit identifier is used to mark which specific simulation time period and time unit these behaviors and results occurred, thus forming a timestamped learning trajectory. For example, in a financial simulation, when a learner agent representing a growth investor completes a valuation analysis of a technology company, its process behavior data (such as consulting industry reports and adjusting growth assumptions) and learning result data (such as the final valuation result and confidence interval) can be stored in its long-term memory along with the identifier of its current "third simulation time period - second simulation time unit".

[0056] S403, during the operation of the simulated teaching cycle, based on the current learning task topic, historical data related to the current learning task topic is retrieved from the long-term memory bank, and the historical data is provided as context information to the corresponding learner agent.

[0057] Here, the learning task theme refers to the core knowledge domain or problem type of the task issued by the teacher agent in the current stage. When the learner agent is detected to have started processing a new task, historical behavior records and result data highly related to the current theme can be retrieved from its long-term memory based on semantic matching or keyword extraction technology. This data is then dynamically inserted into the agent's current processing flow as contextual information to aid decision-making. For example, in a subsequent cycle of a medical simulation, when the learner agent encounters a case involving cardiopulmonary function assessment again, its long-term memory can be automatically retrieved to show its diagnostic process, mistakes made, and related summary notes in handling similar cases in the past. This information can then be pushed to the agent as a reference background for the current diagnosis.

[0058] In this way, by constructing and utilizing a long-term memory bank to achieve the continuous accumulation and contextual retrieval of historical experience, it is possible to simulate advanced cognitive abilities based on experiential reflection and knowledge transfer in actual learning, thereby realizing the coherence modeling of individual learning paths and in-depth analysis of cross-cycle cognitive evolution.

[0059] In some possible embodiments, to more comprehensively simulate the metacognitive and psychological state changes of each learner agent during the learning process and to quantitatively track the evolution of learners' self-confidence, a periodic self-assessment mechanism can be introduced, referring to... Figure 5 As shown, when running the simulated teaching cycle and recording process behavior data, the following steps S501~S503 may also be included: S501, during the simulated teaching cycle, a self-assessment instruction is triggered periodically.

[0060] Understandably, a self-assessment instruction is a request issued by the system that requires the learner agent to reflect on and quantitatively evaluate its current learning status and ability level. It can be automatically triggered by the system according to a preset schedule after completing a key teaching stage (such as a monthly test or a major debate task), aiming to simulate the moment when learners reflect on their own abilities after experiencing a period of learning.

[0061] S502, control the learner agent to respond to the self-assessment instruction and generate a self-concept score, the self-concept score being used to characterize the learner agent's confidence level in its own learning ability.

[0062] Here, the self-concept score is a quantitative value generated by the learner agent through introspection of its recent performance, knowledge mastery, and task completion. For example, in a financial simulation, an analyst agent who has just completed a stress test on a complex portfolio may generate a score reflecting its current confidence (e.g., 80 points, indicating strong confidence) based on its perception of its own model understanding depth and test results. In a medical simulation, a doctor agent who makes a mistake on a rare disease diagnosis trap question may generate a lower self-concept score (e.g., 45 points), reflecting a wavering confidence in its judgment in a specific domain.

[0063] S503, the self-concept score is recorded as part of the process behavior data, and when generating the cognitive strategy diagnostic report, a diagnostic conclusion is generated based on the correlation between the trend of the self-concept score with the simulation time unit and the learning result data.

[0064] Specifically, after collecting and recording self-concept ratings at a series of time points, these can be constructed into time-series data in the order of simulated time units. This data can then be compared with objective learning outcome data (such as test accuracy, trap question recognition rate, and debate performance scores) within the same time period through correlation analysis and pattern comparison. Based on the correlation between the trend of self-concept ratings over simulated time units and the learning outcome data, the learner's metacognitive calibration level and psychological adjustment patterns can be revealed. Qualitative conclusions can then be generated according to pre-defined diagnostic rules. For example, if a learner's rating steadily increases with the improvement of learning outcome data, it may indicate that they possess good self-monitoring and calibration abilities; if their rating is consistently inflated, but their learning outcome data (especially on trap questions) is poor, it may indicate a cognitive risk of "overconfidence" or "self-cognitive bias"; conversely, if the rating remains consistently low, even if the objective results are acceptable, it may suggest a tendency towards "lack of confidence" or "learning anxiety."

[0065] Thus, by introducing a periodic self-assessment mechanism and recording and analyzing the dynamic evolution of self-concept scores, we can quantitatively track the learner's confidence fluctuations and self-cognition changes during the long-term learning process. This enables in-depth diagnosis and revelation of their metacognitive calibration level, psychological adjustment patterns, and potential cognitive biases (such as overconfidence or lack of confidence). In addition to traditional behavior and outcome analysis, this adds modeling and evaluation of the key dimension of the learner's psychological state, making the generated cognitive strategy diagnostic report more comprehensive.

[0066] Furthermore, after obtaining the process behavior data and learning outcome data of the learner agents during the simulated teaching cycle, in-depth computational analysis and pattern mining can be performed to generate a cognitive strategy diagnostic report. This report is an analytical document generated through calculation and comparison based on the process and outcome data recorded by each learner agent, revealing the internal cognitive patterns of each agent. For example, in the financial field, if a learner agent, while learning portfolio theory, shows over-reliance on a single investment strategy in its process behavior data, and its learning outcome data indicates an unsatisfactory risk-reward ratio in its portfolio, the diagnostic report would point out a problem with the learner agent's investment strategy diversity and recommend expanding its investment strategies to build a more reasonable portfolio. Alternatively, in the medical field, if a learner agent, during disease diagnosis learning, spends a significant amount of time researching non-critical symptoms in its process behavior data, and its learning outcome data indicates an inaccurate grasp of key diagnostic indicators, the diagnostic report would point out deficiencies in information filtering and key point grasp, and recommend adjusting its learning strategy to prioritize key diagnostic indicators.

[0067] Specifically, generating a cognitive strategy diagnostic report may also include identifying cognitive characteristics of learners, such as whether they rely on "shortcut learning" or "pattern matching," referring to... Figure 6 As shown, in order to achieve an automated and computable assessment of cognitive depth, the following steps S601~S603 may be included: S601, for each learner agent, obtain the learner agent's answers to the test task and the trap question within the same simulated time unit.

[0068] Understandably, trap questions are those whose semantic logic doesn't align with the taught knowledge, but whose structure is highly similar to the testing task. For example, in valuation learning in finance, the testing task might be calculating the expected return of a stock using the Capital Asset Pricing Model (CAPM) under normal market liquidity conditions; a trap question could be set up in a special scenario where the surface parameters are the same, but the market experiences severe liquidity shortage, requiring the CAPM to be applied again for calculation, in which case the model's core assumptions (perfect market efficiency, ample liquidity) no longer hold. In diagnostic learning in medicine, the testing task might be diagnosing community-acquired pneumonia based on typical symptoms (fever, cough); a trap question could be set up where a patient presents with the same symptoms, but the background information hides that they have recently used a specific antibiotic that might cause "drug fever."

[0069] S602, compare the learner agent's answers to the test task and the trap question.

[0070] Here, in comparing the answers to the test task and the trap questions, we not only compare the correctness of the answers, but also conduct in-depth analysis of the abstract logic of the problem-solving paths and the pattern structures they rely on. We can analyze whether the core formulas, algorithm steps, and key decision variables used in the answers are completely consistent; analyze whether the answers to the trap questions include steps for identifying and responding to changing preconditions, such as whether liquidity risk premiums or drug side effects screening are mentioned; and analyze whether the reasoning chain is closely dependent on the deep semantics of the question or merely matches the surface expression structure of the question (such as keywords and data arrangement order).

[0071] S603, if the comparison results show that, under the different semantic logic, the learner agent gives answers to the test task and the trap question based on the same question structure pattern, then the learner agent's cognitive strategy diagnostic report will indicate the existence of a cognitive tendency of dependent structure matching.

[0072] Here, if the analysis shows that the learner agent's solution to a trap question exhibits high isomorphism with the test task's answer in terms of core steps, formula selection, and parameter application, without any adaptive adjustments to semantic logic changes, then it can be determined that the answer is based on the same problem structure pattern, rather than on a deep understanding and flexible application of the inherent logic of knowledge. For example, in a financial case, if the agent still directly applies the standard CAPM formula to a trap question without considering the impact of liquidity shortages on risk premiums; or in a medical case, if it still directly provides an antibiotic regimen for typical pneumonia without incorporating "drug fever" into the differential diagnosis, then a clear diagnostic label can be generated, indicating that the agent tends to solve problems by matching surface structures, and there is a risk of cognitive fragility and insufficient generalization ability.

[0073] This disclosure introduces a structured comparative diagnostic mechanism using trap questions as a tool, which enables the identification of learners' deep cognitive characteristics of relying on "shortcut learning" or "pattern matching." This expands the assessment dimension from the traditional correctness of answers to the judgment of problem-solving logic and the depth of knowledge application, thereby improving the reliability and efficiency of early detection and diagnosis of individual professional cognitive ability deficiencies in complex and dynamic real business scenarios (such as financial risk judgment and medical clinical decision-making).

[0074] S204, based on the policy characteristics indicated by the cognitive policy evaluation reports of each learner agent, the multiple learner agents are combined to form a multi-agent cluster capable of collaboratively solving the preset scenario problem.

[0075] Here, strategy characteristics refer to the stable, inherent behavioral patterns that agents exhibit when solving problems, as revealed in the cognitive strategy evaluation reports of each learner agent. For example, deep learner types usually correspond to deep analysis and logical deduction characteristics; surface learner types mostly correspond to rapid pattern recognition and process dependence characteristics; lazy learner types may exhibit risk avoidance and seeking the lowest energy consumption path characteristics; and general learner types may show strategic flexibility or serve as a baseline reference.

[0076] Here, after obtaining the policy characteristics from the cognitive policy evaluation reports of each learner agent, agents with different policy characteristics can be configured to specific task nodes according to the problem-solving requirements of the preset scenario, based on the principles of functional complementarity and role collaboration. Rules for information interaction and decision coordination are established for them, thereby forming an organic whole, namely a multi-agent cluster. This cluster can cope with complex problems that are difficult for a single agent to handle by dividing labor and cooperating among its internal agents based on their policy strengths.

[0077] In some possible embodiments, taking a multidisciplinary consultation simulation in the medical field as a preset scenario, based on the cognitive strategy evaluation reports of each learner agent, a deep learner agent that demonstrates solid mechanistic exploration capabilities in the simulation can be configured as the core diagnostic reasoning role, responsible for deducing complex cases from the perspective of pathophysiological mechanisms; a surface learner agent that excels at quickly matching typical signs with disease atlases can be configured as an auxiliary examination interpretation role, responsible for quickly screening key positive findings in images and test reports; and a lazy learner agent that tends to follow clear checklists and avoid complex inferences in the simulation can be configured as a medical standard and checklist verification role, responsible for ensuring that treatment recommendations comply with clinical guidelines and verifying contraindications. In addition, a general learner agent can be used as an information coordinator or redundant backup. The cluster's workflow can be specifically designed as follows: the auxiliary examination interpretation role submits key findings first; the core diagnostic reasoning role conducts in-depth analysis based on these findings and medical history, proposing preliminary diagnostic hypotheses; and the medical standard verification role verifies the compliance and safety of the treatment plan based on the diagnostic hypotheses.

[0078] In some possible implementations, taking dynamic asset allocation review in the financial field as a preset scenario, a deep learner agent that demonstrates deep chain reasoning about macroeconomic cycles in simulations can be configured as a macro strategy analyst; a surface learner agent that excels at quickly identifying similar market technical patterns from historical data can be configured as a market situation awareness agent; and a lazy learner agent that selects low-volatility, high-certainty paths in decision-making simulations can be configured as a risk and compliance review agent; a general learner agent can serve as a data platform or undertake some general analysis. The cluster's collaborative process can be as follows: the market situation awareness agent identifies technical patterns similar to historical "policy bottoms" in real time and issues signals; the macro strategy analyst agent deeply analyzes current economic data and policy texts to assess the strength and sustainability of the macroeconomic logic behind the signals; and the risk and compliance review agent, based on the aforementioned analysis, calculates the specific asset weight adjustment range strictly according to the portfolio's risk budget and compliance constraints. Thus, the cluster can collaboratively complete the entire decision-making chain from market signal capture and macroeconomic logic deepening to risk control implementation, and its efficiency is rooted in the effective utilization and combination of the strategic characteristics of different types of agents.

[0079] The multi-agent cognitive strategy training and evaluation method, system, device and medium provided in this disclosure can simulate the dynamic evolution process of cognitive strategies of agents in long-term, multi-stage tasks by constructing a multi-agent collaborative architecture, thereby achieving a refined evaluation of the cognitive strategies of agents; and based on the policy characteristics obtained from the evaluation, agents with complementary strategies are combined into a collaborative cluster, thereby constructing a multi-agent cluster that can efficiently and robustly solve problems in various scenarios.

[0080] Those skilled in the art will understand that, in the above-described method of the specific implementation, the order in which each step is written does not imply a strict execution order and does not constitute any limitation on the implementation process. The specific execution order of each step should be determined by its function and possible internal logic.

[0081] Based on the same inventive concept, this disclosure also provides a multi-agent cognitive strategy training and evaluation system corresponding to the multi-agent cognitive strategy training and evaluation method. Since the principle of the system in this disclosure for solving the problem is similar to the multi-agent cognitive strategy training and evaluation method described above in this disclosure, the implementation of the system can refer to the implementation of the method, and the repeated parts will not be described again.

[0082] Reference Figure 7 The diagram shown is a schematic of a multi-agent cognitive strategy training and evaluation system provided in an embodiment of this disclosure. The system includes: The agent building unit is used to build and initialize a teacher agent and multiple learner agents with different learning behavior characteristics, including deep learners, surface learners, lazy learners and general learners; The cycle simulation unit is used to run a simulated teaching cycle, control the teacher agent to release learning stage information, and control the learner agent to execute the learning stage information and generate interactive responses to the learning stage information; wherein, the learning stage information is determined based on a preset scenario problem, and the interactive response information includes process behavior data and learning result data; The diagnostic report generation unit is used to record the process behavior data and learning outcome data of each learner agent in the simulated teaching cycle; and to generate a cognitive strategy diagnostic report for each learner agent based on the process behavior data and the outcome data. The agent collaboration construction unit is used to combine the multiple learner agents based on the policy features indicated by the cognitive policy evaluation reports of each learner agent, so as to form a multi-agent cluster capable of collaboratively solving the preset scenario problem.

[0083] In some possible embodiments, the periodic simulation unit is specifically used for: The simulated teaching cycle is divided into multiple consecutive simulated time periods; wherein each simulated time period represents a complete learning cycle and contains multiple consecutive simulated time units executed sequentially. In the first simulation time unit, the teacher agent is controlled to publish knowledge teaching content, and each learner agent is controlled to receive the knowledge teaching content. In the second simulation time unit, the teacher agent is controlled to issue the learning task corresponding to the knowledge teaching content, and each learner agent is controlled to execute the learning task. In the third simulation time unit, the teacher agent is controlled to issue test tasks and trap questions, receive the answers of each learner agent to the test tasks and trap questions, and issue feedback information on the correctness of the answers to each learner agent based on the preset answers. In the fourth simulation time unit, the teacher agent is controlled to organize the learner agents to participate in a debate task on a preset topic, and the speech content and opinion changes of each learner agent in the debate task are recorded.

[0084] In some possible embodiments, the periodic simulation unit is further used for: Each learner agent is sent a strategy selection instruction; wherein the strategy selection instruction provides a summary option and a rest option. Receive the selection results returned by each learner agent based on its configured type; If the selected result is the summary option, a learning summary record is generated and stored for the learner agent. If the selected result is the rest option, then a rest behavior is recorded for the learner agent, and the step of generating a learning summary record is skipped.

[0085] In some possible embodiments, the diagnostic report generation unit is specifically used for: For each learner agent, obtain the learner agent's answers to the test task and the trap question within the same simulated time unit; wherein, the trap question is a question whose semantic logic does not match the knowledge taught content, but whose question structure is similar to the test task; Compare the learner agent's answers to the test task and the trap question; If the comparison results show that, under different semantic logic, the learner agent gives answers to the test task and the trap question based on the same question structure pattern, then the learner agent's cognitive strategy diagnostic report will indicate a cognitive tendency of dependent structure matching.

[0086] In some possible embodiments, the diagnostic report generation unit is specifically used for: Establish and maintain a long-term memory bank for each learner agent; For each learner agent, the process behavior data and learning result data of the learner agent, as well as the simulated time unit identifiers corresponding to each data, are stored in the long-term memory bank corresponding to the learner agent. During the simulated teaching cycle, based on the current learning task topic, historical data related to the current learning task topic is retrieved from the long-term memory, and the historical data is provided as context information to the corresponding learner agent.

[0087] In some possible embodiments, the diagnostic report generation unit is specifically used for: During the simulated teaching cycle, self-assessment instructions are triggered periodically; The learner agent is controlled to respond to the self-assessment instruction and generate a self-concept score, which is used to characterize the learner agent's confidence level in its own learning ability. The self-concept score is recorded as part of the process behavior data, and when generating the cognitive strategy diagnostic report, a diagnostic conclusion is generated based on the correlation between the trend of the self-concept score changing with the simulated time unit and the learning outcome data.

[0088] Based on the same inventive concept, this disclosure also provides a multi-agent cognitive strategy training and evaluation device corresponding to the multi-agent cognitive strategy training and evaluation method. Since the principle of the device in this disclosure for solving the problem is similar to the multi-agent cognitive strategy training and evaluation method described above in this disclosure, the implementation of the device can refer to the implementation of the method, and the repeated parts will not be described again.

[0089] Reference Figure 8 The diagram shown is a schematic of a multi-agent cognitive strategy training and evaluation device 800 provided in an embodiment of this disclosure. The device includes: The agent construction module 801 is used to construct and initialize a teacher agent and multiple learner agents; wherein the learner agents are configured as types with different learning behavior characteristics, including deep learners, surface learners, lazy learners and general learners; The teaching simulation module 802 is used to run a simulated teaching cycle, in which the teacher agent is controlled to publish learning stage information to all learner agents; and the learner agents are controlled to execute the learning stage information based on their respective configured types, and generate interactive response information for the learning stage information; wherein the learning stage information is determined based on a preset scenario problem, and the interactive response information includes process behavior data and learning result data; The report generation module 803 is used to record the process behavior data and learning outcome data of each learner agent in the simulated teaching cycle; and to generate a cognitive strategy diagnostic report for each learner agent based on the process behavior data and the outcome data. The agent combination module 804 is used to combine the multiple learner agents based on the policy characteristics indicated by the cognitive policy evaluation reports of each learner agent, so as to form a multi-agent cluster that can collaboratively solve the preset scenario problem.

[0090] In some possible embodiments, the teaching simulation module 802 is specifically used for: The simulated teaching cycle is divided into multiple consecutive simulated time periods; wherein each simulated time period represents a complete learning cycle and contains multiple consecutive simulated time units executed sequentially. In the first simulation time unit, the teacher agent is controlled to publish knowledge teaching content, and each learner agent is controlled to receive the knowledge teaching content. In the second simulation time unit, the teacher agent is controlled to issue the learning task corresponding to the knowledge teaching content, and each learner agent is controlled to execute the learning task. In the third simulation time unit, the teacher agent is controlled to issue test tasks and trap questions, receive the answers of each learner agent to the test tasks and trap questions, and issue feedback information on the correctness of the answers to each learner agent based on the preset answers. In the fourth simulation time unit, the teacher agent organizes the learner agents to participate in a debate task on a preset topic, and records the speech content and changes in viewpoints of each learner agent during the debate task. In some possible embodiments, the teaching simulation module 802 is further configured to: Each learner agent is sent a strategy selection instruction; wherein the strategy selection instruction provides a summary option and a rest option. Receive the selection results returned by each learner agent based on its configured type; If the selected result is the summary option, a learning summary record is generated and stored for the learner agent. If the selected result is the rest option, then a rest behavior is recorded for the learner agent, and the step of generating a learning summary record is skipped.

[0091] In some possible embodiments, the report generation module 803 is specifically used for: For each learner agent, obtain the learner agent's answers to the test task and the trap question within the same simulated time unit; wherein, the trap question is a question whose semantic logic does not match the knowledge taught content, but whose question structure is similar to the test task; Compare the learner agent's answers to the test task and the trap question; If the comparison results show that, under different semantic logic, the learner agent gives answers to the test task and the trap question based on the same question structure pattern, then the learner agent's cognitive strategy diagnostic report will indicate a cognitive tendency of dependent structure matching.

[0092] In some possible embodiments, the report generation module 803 is specifically used for: Establish and maintain a long-term memory bank for each learner agent; For each learner agent, the process behavior data and learning result data of the learner agent, as well as the simulated time unit identifiers corresponding to each data, are stored in the long-term memory bank corresponding to the learner agent. During the simulated teaching cycle, based on the current learning task topic, historical data related to the current learning task topic is retrieved from the long-term memory, and the historical data is provided as context information to the corresponding learner agent.

[0093] In some possible embodiments, the report generation module 803 is specifically used for: During the simulated teaching cycle, self-assessment instructions are triggered periodically; The learner agent is controlled to respond to the self-assessment instruction and generate a self-concept score, which is used to characterize the learner agent's confidence level in its own learning ability. The self-concept score is recorded as part of the process behavior data, and when generating the cognitive strategy diagnostic report, a diagnostic conclusion is generated based on the correlation between the trend of the self-concept score changing with the simulated time unit and the learning outcome data.

[0094] Based on the same technical concept, this disclosure also provides a computer device. (See also...) Figure 9The diagram shown is a structural schematic of a computer device 900 provided in an embodiment of this disclosure, including a processor 901, a memory 902, and a bus 903. The memory 902 stores execution instructions and includes a main memory 9021 and an external memory 9022. The main memory 9021, also called internal memory, is used to temporarily store computational data in the processor 901, as well as data exchanged with external memory 9022 such as a hard disk. The processor 901 exchanges data with the external memory 9022 through the main memory 9021.

[0095] In this embodiment, the memory 902 is specifically used to store application code that executes the solution of this application, and its execution is controlled by the processor 901. That is, when the computer device 900 is running, the processor 901 communicates with the memory 902 through the bus 903, so that the processor 901 executes the application code stored in the memory 902, and then executes the method described in any of the foregoing embodiments.

[0096] The memory 902 may be, but is not limited to, random access memory (RAM), read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), etc.

[0097] Processor 901 may be an integrated circuit chip with signal processing capabilities. The aforementioned processor can be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; it can also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this invention. The general-purpose processor can be a microprocessor or any conventional processor.

[0098] It is understood that the structures illustrated in the embodiments of this application do not constitute a specific limitation on the computer device 900. In other embodiments of this application, the computer device 900 may include more or fewer components than illustrated, or combine some components, or split some components, or have different component arrangements. The illustrated components may be implemented in hardware, software, or a combination of software and hardware.

[0099] This disclosure also provides a computer-readable storage medium storing a computer program. When executed by a processor, the computer program performs the steps of the multi-agent cognitive strategy training and evaluation method described in the above-described method embodiments. The storage medium can be volatile or non-volatile computer-readable storage.

[0100] This disclosure also provides a computer program product carrying program code. The program code includes instructions that can be used to execute the steps of the multi-agent cognitive strategy training and evaluation method described in the above method embodiments. For details, please refer to the above method embodiments, which will not be repeated here.

[0101] The aforementioned computer program product can be implemented through hardware, software, or a combination thereof. In one optional embodiment, the computer program product is specifically embodied in a computer storage medium; in another optional embodiment, the computer program product is specifically embodied in a software product, such as a software development kit (SDK), etc.

[0102] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems and devices described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here. In the several embodiments provided in this disclosure, it should be understood that the disclosed systems and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division; in actual implementation, there may be other division methods. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Another point is that the displayed or discussed mutual coupling or direct coupling or communication connection may be through some communication interfaces; the indirect coupling or communication connection of devices or units may be electrical, mechanical, or other forms.

[0103] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0104] In addition, the functional units in the various embodiments of this disclosure can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0105] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a processor-executable, non-volatile, computer-readable storage medium. Based on this understanding, the technical solution of this disclosure, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this disclosure. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0106] Finally, it should be noted that the above-described embodiments are merely specific implementations of this disclosure, used to illustrate the technical solutions of this disclosure, and not to limit it. The protection scope of this disclosure is not limited thereto. Although this disclosure has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some of the technical features, within the scope of the technology disclosed in this disclosure. Such modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this disclosure, and should all be covered within the protection scope of this disclosure. Therefore, the protection scope of this disclosure should be determined by the protection scope of the claims.

Claims

1. A method for multi-agent cognitive policy training and evaluation, characterized in that, include: Construct and initialize a teacher agent and multiple learner agents; wherein the learner agents are configured as types with different learning behavior characteristics, including deep learners, surface learners, lazy learners and general learners; A simulated teaching cycle is run, in which the teacher agent is controlled to publish learning stage information to all learner agents; and the learner agents are controlled to execute the learning stage information based on their respective configured types, and generate interactive response information in response to the learning stage information; wherein the learning stage information is determined based on a preset scenario problem, and the interactive response information includes process behavior data and learning result data; Record the process behavior data and learning outcome data of each learner agent during the simulated teaching cycle; and generate a cognitive strategy diagnostic report for each learner agent based on the process behavior data and the outcome data. Based on the policy characteristics indicated by the cognitive policy evaluation reports of each learner agent, the multiple learner agents are combined to form a multi-agent cluster capable of collaboratively solving the preset scenario problem.

2. The method according to claim 1, characterized in that, Running a simulated teaching cycle includes: The simulated teaching cycle is divided into multiple consecutive simulated time periods; wherein each simulated time period represents a complete learning cycle and contains multiple consecutive simulated time units executed sequentially. In the first simulation time unit, the teacher agent is controlled to publish knowledge teaching content, and each learner agent is controlled to receive the knowledge teaching content. In the second simulation time unit, the teacher agent is controlled to issue the learning task corresponding to the knowledge teaching content, and each learner agent is controlled to execute the learning task. In the third simulation time unit, the teacher agent is controlled to issue test tasks and trap questions, receive the answers of each learner agent to the test tasks and trap questions, and issue feedback information on the correctness of the answers to each learner agent based on the preset answers. In the fourth simulation time unit, the teacher agent is controlled to organize the learner agents to participate in a debate task on a preset topic, and the speech content and opinion changes of each learner agent in the debate task are recorded.

3. The method according to claim 2, characterized in that, Following the third simulation time unit, the following is included: Each learner agent is sent a strategy selection instruction; wherein the strategy selection instruction provides a summary option and a rest option. Receive the selection results returned by each learner agent based on its configured type; If the selected result is the summary option, a learning summary record is generated and stored for the learner agent. If the selected result is the rest option, then a rest behavior is recorded for the learner agent, and the step of generating a learning summary record is skipped.

4. The method of claim 3, wherein, The generation of a cognitive strategy diagnostic report for each learner agent includes: For each learner agent, obtain the learner agent's answers to the test task and the trap question within the same simulated time unit; wherein, the trap question is a question whose semantic logic does not match the knowledge taught content, but whose question structure is similar to the test task; Compare the learner agent's answers to the test task and the trap question; If the comparison results show that, under different semantic logic, the learner agent gives answers to the test task and the trap question based on the same question structure pattern, then the learner agent's cognitive strategy diagnostic report will indicate a cognitive tendency of dependent structure matching.

5. The method of claim 1, wherein, The recorded process behavior data and learning outcome data of each learner agent during the simulated teaching cycle include: Establish and maintain a long-term memory bank for each learner agent; For each learner agent, the process behavior data and learning result data of the learner agent, as well as the simulated time unit identifiers corresponding to each data, are stored in the long-term memory bank corresponding to the learner agent. During the simulated teaching cycle, based on the current learning task topic, historical data related to the current learning task topic is retrieved from the long-term memory, and the historical data is provided as context information to the corresponding learner agent.

6. The method of claim 3, wherein, The recorded process behavior data and learning outcome data of each learner agent during the simulated teaching cycle include: During the simulated teaching cycle, self-assessment instructions are triggered periodically; The learner agent is controlled to respond to the self-assessment instruction and generate a self-concept score, which is used to characterize the learner agent's confidence level in its own learning ability. The self-concept score is recorded as part of the process behavior data, and when generating the cognitive strategy diagnostic report, a diagnostic conclusion is generated based on the correlation between the trend of the self-concept score changing with the simulated time unit and the learning outcome data.

7. A multi-agent cognitive policy training and evaluation system, comprising: include: The agent building unit is used to build and initialize a teacher agent and multiple learner agents with different learning behavior characteristics, including deep learners, surface learners, lazy learners and general learners; The cycle simulation unit is used to run a simulated teaching cycle, control the teacher agent to release learning stage information, and control the learner agent to execute the learning stage information and generate interactive responses to the learning stage information; wherein, the learning stage information is determined based on a preset scenario problem, and the interactive response information includes process behavior data and learning result data; The diagnostic report generation unit is used to record the process behavior data and learning outcome data of each learner agent in the simulated teaching cycle; and to generate a cognitive strategy diagnostic report for each learner agent based on the process behavior data and the outcome data. The agent collaboration construction unit is used to combine the multiple learner agents based on the policy features indicated by the cognitive policy evaluation reports of each learner agent, so as to form a multi-agent cluster capable of collaboratively solving the preset scenario problem.

8. A multi-agent cognitive policy training and evaluation apparatus, characterized in that, include: The agent construction module is used to construct and initialize a teacher agent and multiple learner agents; wherein, the learner agents are configured as types with different learning behavior characteristics, including deep learners, surface learners, lazy learners and general learners; The teaching simulation module is used to run a simulated teaching cycle. During the simulated teaching cycle, the teacher agent is controlled to publish learning stage information to all learner agents. The learner agents are also controlled to execute the learning stage information based on their respective configured types and generate interactive response information in response to the learning stage information. The learning stage information is determined based on a preset scenario problem, and the interactive response information includes process behavior data and learning result data. The report generation module is used to record the process behavior data and learning outcome data of each learner agent in the simulated teaching cycle; and to generate a cognitive strategy diagnostic report for each learner agent based on the process behavior data and the outcome data. The agent combination module is used to combine the multiple learner agents based on the policy characteristics indicated by the cognitive policy evaluation reports of each learner agent, so as to form a multi-agent cluster that can collaboratively solve the preset scenario problem.

9. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 6.

10. A computer device comprising a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 6.

Images