Multi-agent cognitive training game generation system for cognitive impairment

Abstract

The application discloses a kind of multi-agent cognitive training game generation systems for cognitive impairment, belong to cognitive training technical field.The application first obtains the portrait parameter of target patient, then extracts the historical evidence data of each candidate collaborative configuration from configuration evidence base, according to this and portrait parameter determine adaptive cost penalty weight and evidence sparsity penalty weight, then calculate the utility value of each candidate collaborative configuration and determine target collaborative configuration data according to preset determination logic, subsequently activate corresponding agent role according to the configuration, generate initial game design data and calculate consensus score in multiple rounds of interaction of portrait parameter, finally generate cognitive training game according to this and collect answer behavior dynamic adjustment difficulty and feedback, meet the personalized needs of patient population cognitive training, solve the problem that existing system is difficult to meet the personalized needs of patient population cognitive training.

Description

Technical Field

[0001] This invention relates to the field of cognitive training technology, and in particular to a multi-agent cognitive training game generation system for cognitive impairment. Background Technology

[0002] With the accelerating aging of the population, the number of people suffering from cognitive degenerative diseases, such as Alzheimer's disease and mild cognitive impairment, continues to rise, making the need for cognitive rehabilitation training for this population increasingly urgent. Cognitive training aims to intervene in different cognitive domains, such as memory, attention, and executive function, through targeted tasks. The training content needs to be personalized according to the patient's age, disease stage, and the degree of impairment in each cognitive domain, thus placing high demands on the professionalism and adaptability of the training tasks.

[0003] To alleviate the aforementioned supply pressure, content generation technology based on large language models has been introduced into the field of automatic generation of cognitive training tasks in recent years. In practice, to improve the professionalism and reliability of the generated content, the industry has further adopted a multi-agent collaborative approach. This involves multiple agents, each with different functions such as requirements analysis, solution design, content implementation, and quality review, collaborating with each other to complete the generation of training tasks through role division and multiple rounds of interaction.

[0004] However, existing multi-agent collaborative generation methods generally employ a fixed collaborative architecture, meaning that regardless of the patient or cognitive domain, all agent roles are activated and the entire collaborative interaction process is executed. In this approach, each task generation requires calling all participating roles and completing multiple rounds of interaction, resulting in significant model calls and computational overhead. In reality, the required collaboration intensity to achieve satisfactory generation quality varies depending on different patient conditions and cognitive domains. In some cases, activating all roles or executing all interaction rounds is sufficient to obtain adequate generation quality. The fixed architecture fails to differentiate the level of participation based on specific patient conditions, causing a large amount of computational resources to be consumed on redundant interactions that do not substantially improve generation quality.

[0005] Therefore, the existing cognitive training task generation technology based on multi-agent collaboration has a fixed collaborative architecture and cannot adjust the collaboration intensity according to the patient's condition. This makes it impossible for the system to be personalized according to the cognitive profile and disease stage of an individual patient, making it difficult to meet the personalized needs of a large number of patients. Summary of the Invention

[0006] To address the technical problem that existing systems cannot be personalized based on individual patient cognitive profiles and disease stages, thus failing to meet the personalized needs of large-scale patient populations, this invention provides a multi-agent cognitive training game generation system for cognitive impairment. The technical solution is as follows:

[0007] The profile parameter acquisition module is used to acquire profile parameters of the target patient, including the target training domain and cognitive ability score.

[0008] The calibration perception routing module receives profile parameters and extracts historical evidence data of each candidate collaborative configuration from a pre-built configuration evidence library. It determines adaptive cost penalty weights based on the target training domain, calculates evidence sparsity penalty weights based on the cumulative number of evidence items in the historical evidence data, and calculates the utility value of each candidate collaborative configuration by combining historical evidence data and cost penalty weights. It processes the utility value according to the preset judgment logic and determines the target collaborative configuration data, where the candidate collaborative configuration is the combination of the on or off states of multiple intelligent agent roles in the system.

[0009] A configurable multi-agent negotiation module is used to acquire target collaboration configuration data, activate the corresponding agent role according to the target collaboration configuration data, perform multiple rounds of interactive processing on the profile parameters, generate initial game design data, and calculate the corresponding consensus score.

[0010] The game execution and adaptive intervention module is used to receive initial game design data to generate cognitive training game data, collect game answering behavior data of target patients when executing cognitive training game data to dynamically calculate difficulty adjustment parameters and feed them back to the subsequent game generation process, and package the generated resource consumption data, activation component data and consensus score into audit logs and send them to the configuration evidence library for updating.

[0011] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following:

[0012] 1. This invention provides a multi-agent cognitive training game generation system for cognitive impairment. Through the coordinated operation of a profile parameter acquisition module, a calibration perception routing module, a configurable multi-agent negotiation module, and a game execution and adaptive intervention module, the multi-agent collaboration is transformed from a fixed, unchanging full-role collaboration architecture to an adjustable architecture that can be adjusted according to the patient's conditions. Furthermore, this solution uses the profile parameters of the target patient as the decision-making basis. Before executing the collaboration generation, a target collaboration configuration that achieves a balance between generation quality and resource consumption is selected from the candidate collaboration configurations. Based on this, the corresponding agent roles are selectively activated to complete the game generation. The evidence of each generation is written back to the configuration evidence library to continuously calibrate the routing decision. This not only reduces the computational resource consumption of the collaboration process while ensuring the generation quality, but also meets the personalized needs of cognitive training for the patient population, realizing the automatic generation of cognitive training games with a dynamic balance between resource consumption and generation quality.

[0013] 2. The calibration-aware routing module of the present invention uniformly measures each candidate cooperative configuration with a utility value. The cost penalty weight is adaptive with the fault tolerance of the target training domain, and the evidence sparsity penalty weight is adaptive with the number of accumulated evidence. This allows the routing to balance quality and resources under different cognitive domains and different levels of evidence sufficiency. Furthermore, with the help of the hysteresis determination submodule, the configuration is switched only when the utility improvement exceeds the hysteresis threshold. This effectively suppresses repeated switching and regeneration caused by small fluctuations in utility between adjacent rounds, reduces resource consumption fluctuations, and improves system operation stability.

[0014] 3. The early calibration module of this invention performs sequential simulation evaluation on the minimum calibration configuration set before calculating the utility value, and terminates the evaluation in advance when the leading advantage of the current optimal configuration exceeds the early stop threshold, so that resource conservation extends to the evidence collection stage; the cold start rollback submodule rolls back step by step according to cognitive domain level evidence and behavioral prototype patient group evidence when there is a lack of evidence for the target patient, and adds confidence level labels, so that there is still evidence-based routing decision in the new patient scenario, improving the routing availability and decision reliability of the system in the cold start scenario. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 A schematic diagram of the structure of a multi-agent cognitive training game generation system for cognitive impairment provided in this application embodiment;

[0017] Figure 2A flowchart illustrating the tripartite negotiation process provided in this application embodiment. Detailed Implementation

[0018] Embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While some embodiments of the present disclosure are shown in the drawings, it should be understood that embodiments of the present disclosure may be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of the present disclosure.

[0019] It should be understood that the accompanying drawings and embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of protection of this disclosure. In the description of the embodiments of this disclosure, the term "comprising" and similar terms should be understood as open-ended inclusion, i.e., "including but not limited to". The term "based on" should be understood as "at least partially based on". The term "one embodiment" or "this embodiment" should be understood as "at least one embodiment". The terms "first", "second", etc., may refer to different or the same objects.

[0020] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0021] like Figure 1 The diagram shown is a structural schematic of a multi-agent cognitive training game generation system for cognitive impairment provided in an embodiment of this application. In implementation, the system is deployed on a server or edge computing device with computing capabilities and communicates with an external diagnostic system storing profile parameters, a configuration evidence library storing historical evidence data, and a model service providing large language model reasoning capabilities. The system includes a profile parameter acquisition module, a calibration perception routing module, a configurable multi-agent negotiation module, and a game execution and adaptive intervention module, all connected internally. It may further include an early calibration module and a behavioral simulation modeling module communicating with a grouped database.

[0022] In this system, each intelligent agent role, responsible for functions such as requirements analysis, design, content implementation, and quality review, is implemented based on a pre-trained large language model. Each role is assigned corresponding responsibilities and constraints through preset role prompts, thus playing collaborative roles with different divisions of labor on the same underlying model. It should be noted that the requirements analysis submodule can be implemented as a neuropsychologist role responsible for clinical needs analysis and cognitive effectiveness constraints; the design submodule can be implemented as a game designer role responsible for game mechanics and playability design; the review and assembly submodule can be implemented as a quality reviewer role that oversees from multiple dimensions; and the coding submodule can be implemented as a programmer role that completes the specific content implementation. The above role names are only exemplary implementations and do not constitute a limitation on their specific forms. The specific implementations of each module are as follows:

[0023] The profile parameter acquisition module is used to acquire profile parameters of the target patient, including the target training domain and cognitive ability score.

[0024] The target training domain reflects the cognitive domain targeted in this training, while the cognitive ability score reflects the degree of impairment or retention of the target patient in each cognitive domain. It should be noted that the cognitive ability score can be derived from the assessment results of standardized neuropsychological scales, statistical results from the patient's previous training records, or imported from an external diagnostic system via an interface. In addition to the target training domain and cognitive ability score, the profile parameters can further include conditional data such as age and disease stage to improve the accuracy of subsequent routing and generation. The profile parameters output by this module are simultaneously provided to the calibration perception routing module and the configurable multi-agent negotiation module. The former uses these parameters to select collaborative configurations, while the latter uses them to generate game content, ensuring that routing decisions and content generation are based on the same patient conditions.

[0025] The calibration perception routing module receives profile parameters and extracts historical evidence data of each candidate collaborative configuration from a pre-built configuration evidence library. It determines adaptive cost penalty weights based on the target training domain, calculates evidence sparsity penalty weights based on the cumulative number of evidence items in the historical evidence data, and calculates the utility value of each candidate collaborative configuration by combining historical evidence data and cost penalty weights. It processes the utility value according to the preset judgment logic and determines the target collaborative configuration data, where the candidate collaborative configuration is the combination of the on or off states of multiple intelligent agent roles in the system.

[0026] Before generating collaborations, this module selects a target collaboration configuration that balances quality and resource consumption from multiple candidate configurations based on the target patient's conditions. This avoids wasting computational resources by activating all agent roles for all patients. It's important to note that candidate collaboration configurations are combinations of agent roles such as cognitive analysis, design, coding, and quality review, each with their roles enabled or disabled. This results in configurations with varying strengths, such as enabling all roles, disabling the quality review role, disabling the cognitive analysis role, or even retaining only a single role, making collaboration strength a selectable variable. The configuration evidence base pre-stores historical evidence data accumulated during the generation process of each candidate collaboration configuration. This historical evidence data includes at least generation quality data and resource consumption data, gradually written back from audit logs by the game execution and adaptive intervention module. During initial system runtime, the configuration evidence base can be initialized with simulated evaluation evidence collected by the early calibration module. This module selects a configuration by calculating the utility value of each candidate collaboration configuration. The calculation logic is as follows:

[0027] ;

[0028] In the formula, Indicates candidate collaboration configuration The utility value is used as a dependent variable to measure the overall merits of the configuration; Indicating in historical evidence data Configuration under conditions The estimated quality of the generated data; Indicates configuration Resource consumption data is used to characterize the amount of resources required to generate this configuration in one go; This represents the cost penalty weight, whose value depends on the cognitive domain corresponding to the target training domain. Adaptive determination; Indicates configuration The sparsity of evidence is used to characterize the sufficiency of its historical evidence data; This represents the sparsity penalty weight for evidence, and its value varies depending on the configuration. The cumulative number of evidence pieces Confirmed. This formula uses generation quality as a positive benefit and resource consumption and evidence sparsity as two negative penalties, ensuring that the selected configuration guarantees generation quality while also considering resource conservation and decision reliability. After obtaining each utility value, this module selects the candidate collaboration configuration with the highest utility value as the target collaboration configuration data according to preset judgment logic. This allows for the output of matching collaboration configurations for different patient conditions, reducing redundant agent role calls and multi-turn interactions while ensuring generation quality, thus reducing computational resource consumption during the collaboration process.

[0029] It should be noted that resource consumption data can be measured by the number of lexical units consumed by the large language model call generated during the candidate collaboration configuration's generation process, or by indicators such as the number of model calls and generation latency, either individually or in a weighted comprehensive manner. This embodiment does not limit the specific measurement caliber of resource consumption data. Among them, the large language model on which each intelligent agent role relies can be a model that has been pre-trained on massive text corpora and fine-tuned with cognitive training domain corpora. The responsibilities and constraints of each role are configured through preset role prompt words.

[0030] Furthermore, the calibration-aware routing module includes a utility calculation submodule, which is used to specifically implement the above utility value calculation:

[0031] First, historical quality assessment data and historical resource consumption data for each candidate collaboration configuration are extracted from historical evidence data.

[0032] Historical quality assessment data reflects the previous generation quality level achieved by the configuration, corresponding to the generation quality item in the utility value; historical resource consumption data reflects the amount of resources consumed in a single generation of the configuration, corresponding to the resource consumption item in the utility value. Both are accumulated by the game execution and adaptive intervention module through audit log writing, providing a data foundation for subsequent item-by-item calculation of the utility value.

[0033] It should be added that the historical quality assessment data comes from the consensus scores of each previous generation, which were written back to the configuration evidence base through audit logs. In other words, the consensus scores obtained from each generation are used as the generation quality records of the corresponding candidate collaboration configurations in the historical evidence data.

[0034] Next, a positive correlation is established between the fault tolerance of the target training domain and the cost penalty weight, so as to output the corresponding cost penalty weight.

[0035] It's important to explain that fault tolerance reflects the tolerance level of the corresponding cognitive domain for generation quality deviations. The logic behind establishing a positive correlation between fault tolerance and cost penalty weights is as follows: for cognitive domains with higher fault tolerance, even if configurations with lower resource consumption and weaker collaboration are used, the impact of quality deviations on training results is relatively limited. Therefore, a larger cost penalty weight is output to enhance the negative effect of resource consumption, prompting the system to tend to use configurations with lower resource consumption. Conversely, for cognitive domains with lower fault tolerance, a smaller cost penalty weight is output to weaken the negative effect of resource consumption and prioritize quality. Thus, the cost penalty weights adaptively change with the target training domain, achieving differentiated routing for quality-first or resource-first approaches in different cognitive domains. This positive correlation mapping can be implemented using a fault tolerance gradation and weight mapping table, or it can be implemented using a continuous monotonically increasing function.

[0036] It should be noted that the tolerance can be pre-defined and stored in a preset tolerance configuration table by professionals in the relevant fields based on the training characteristics of each cognitive domain. For example, a lower tolerance can be given to cognitive domains where generation quality deviation is likely to cause training to deviate from the clinical goal, while a higher tolerance can be given to cognitive domains where generation quality deviation has a limited impact on training effect. The tolerance can also be obtained by statistically analyzing the historical evidence data accumulated in the configuration evidence base. That is, the tolerance can be calculated by inversely based on the correlation between generation quality deviation and final training feedback in each cognitive domain, and updated periodically as historical data accumulates.

[0037] Then, a negative correlation mapping relationship is established between the cumulative number of evidence items and the evidence sparsity penalty weight, so as to output the corresponding evidence sparsity penalty weight.

[0038] It should be explained that the more evidence accumulated by a candidate collaborative configuration, the more reliably its historical quality assessment data reflects the actual generated quality; therefore, a smaller penalty weight is output. Conversely, when the number of evidence pieces is small, the reliability of the quality estimate is insufficient; therefore, a larger penalty weight is output to prevent the system from rashly selecting a configuration when evidence is insufficient. One feasible form is to divide the baseline weight by the cumulative number of evidence pieces plus one. Alternatively, other functions that monotonically decrease with the cumulative number of evidence pieces or hierarchical comparison methods can be used.

[0039] Finally, the utility value of each candidate collaboration configuration is calculated by subtracting the product of historical resource consumption data and cost penalty weight from historical quality assessment data, and then subtracting the product of the set state penalty coefficient and evidence sparsity penalty weight. The calculation is as follows:

[0040] ;

[0041] In the formula, This represents the historical quality assessment data of the configuration, as a positive benefit item; This indicates the historical resource consumption data for this configuration. This constitutes a resource consumption penalty item; Indicates the sparsity penalty weight of evidence. This represents the set state penalty coefficient. This constitutes a penalty term for evidence sparsity. This calculation integrates three factors—generation quality, resource consumption, and evidence sufficiency—into a single quantitative indicator, ensuring that the final selected configuration balances quality with resource conservation and decision reliability.

[0042] Furthermore, the utility calculation submodule also includes a time-weighted unit, used to correct the time-related aspects of historical quality assessment data before using it for utility value calculation:

[0043] The first step is to obtain the generation timestamp corresponding to each piece of evidence in the historical quality assessment data, and calculate the difference between the generation timestamp and the current system time.

[0044] It should be noted that, considering that both the large language model used and the state of the target patient may change over time, earlier accumulated historical quality assessment data may no longer accurately reflect the current actual generation quality. The generation timestamp is recorded along with each piece of evidence when it is written back to the configuration evidence library through the audit log. The difference data reflects the length of time that the evidence has been since its generation, i.e., the freshness of the evidence, providing a quantitative basis for determining the weight based on freshness in the subsequent process.

[0045] The second step is to calculate the time decay coefficient based on the difference data and construct a negative correlation function between the difference data and the time decay coefficient.

[0046] This negative correlation function assigns a smaller time decay coefficient to larger differences and a larger time decay coefficient to smaller differences, thus giving higher weight to newer evidence and lower weight to older evidence, achieving a gradual forgetting of historical evidence. This function can be implemented using an exponential function with a natural constant as the base and the negative of the ratio of the difference data to a preset decay constant as the exponent, or it can be implemented using other functions that decrease as the difference increases. The rate of decay can be adjusted through corresponding decay parameters.

[0047] The third step involves using the time decay coefficient to perform a weighted summation of each historical quality assessment data point, generating revised quality assessment data. This revised quality assessment data then replaces the original historical quality assessment data in the utility value calculation. The calculation logic is as follows:

[0048] ;

[0049] In the formula, Indicates configuration Corrected quality assessment data; This indicates that the configuration is number 1. Historical quality assessment data; Indicates the first The smaller the difference in the time decay coefficients corresponding to the pieces of evidence, the better. The larger the value, the better. This formula normalizes and weights each historical quality assessment data point by using a time decay coefficient as the weight, making the revised quality assessment data a weighted quality value that emphasizes newer evidence. This allows routing decisions to be adjusted in a timely manner to keep up with model iterations and changes in patient status, thus maintaining the accuracy of quality estimation even when the model or patient status changes.

[0050] Furthermore, the calibration-aware routing module includes a hysteresis determination submodule, used to execute the aforementioned preset determination logic:

[0051] The first step is to compare the calculated utility values ​​and extract the candidate collaboration configuration corresponding to the maximum value as the first collaboration configuration data.

[0052] Among them, the first collaborative configuration data is the candidate configuration with the best utility value under the current patient conditions and evidence status. Whether it is ultimately confirmed as the target collaborative configuration data still needs to be determined by subsequent delay judgment, rather than being directly adopted.

[0053] The second step is to obtain the second collaboration configuration data used in the previous game generation round and extract the reference utility value corresponding to the second collaboration configuration data.

[0054] The second collaborative configuration data is the configuration currently in use by the system. The reference utility value reflects the utility level of the current configuration under the current conditions and serves as a comparison benchmark for judging whether it is worthwhile to switch, so that the switching decision is based on the utility difference between the old and new configurations.

[0055] The third step is to calculate the difference between the utility value of the first collaborative configuration data and the reference utility value.

[0056] The difference data reflects the utility improvement that would result from switching from the current configuration to the optimal configuration, and serves as the basis for subsequent hysteresis determination.

[0057] Fourth step: When the difference data is greater than the preset hysteresis threshold, the first collaboration configuration data is confirmed as the target collaboration configuration data; when the difference data is not greater than the preset hysteresis threshold, the second collaboration configuration data is kept as the target collaboration configuration data.

[0058] It should be noted that the preset hysteresis threshold has a diversion effect on subsequent operations. The larger the value, the higher the utility improvement required to trigger the switch, and the stronger the system's tendency to retain the original configuration; the smaller the value, the more sensitive the system is to changes in utility and the easier it is to switch. Switching is only performed when the difference data is greater than this threshold, meaning the utility improvement is sufficient to offset the overhead of reorganizing collaboration generation due to the configuration switch. Otherwise, the original configuration is maintained, thus forming a stable interval between adjacent rounds, suppressing repeated switching of configurations between different intensities, reducing repeated collaboration generation calls caused by frequent switching, and achieving the effect of reducing resource consumption fluctuations and improving system operational stability. When the difference data is exactly equal to the threshold, it can be categorized as maintaining the original configuration. The threshold can be a fixed value or dynamically set according to the switching overhead or operating status.

[0059] Furthermore, the calibration-aware routing module also includes a cold start rollback submodule, used to perform the following processing when extracting historical evidence data:

[0060] First, check if there is historical evidence data of the target patient in the configured evidence base that perfectly matches the profile parameters. If not, perform a step-by-step query according to priority.

[0061] Among them, the historical evidence data of the target patient is the evidence accumulated by the patient in the previous generation process, which is most consistent with the current patient conditions and is therefore given the highest priority; if it does not exist, it means that the system is still in a cold start state for the patient, and the search is relaxed level by level from high to low according to the degree of consistency with the target patient.

[0062] Next, query the configuration evidence base to see if there is cognitive domain-level evidence that matches the target training domain. If it exists, extract it as the first alternative evidence and attach a first confidence level label.

[0063] Among them, cognitive domain-level evidence is aggregated evidence accumulated by other patients in the same cognitive domain. Although it matches the target training domain, it does not originate from the target patient. Therefore, a first confidence level label is attached to characterize its credibility relative to the target patient, so that subsequent logic can know its credibility level when using it.

[0064] Then, if no cognitive domain-level evidence exists, the profile parameters are sent to a pre-defined group database for matching, the corresponding behavioral prototype patient group data is obtained, the group evidence data is extracted as second alternative evidence, and a second confidence level label is attached.

[0065] Among them, the behavioral prototype patient group data consists of patient groups divided according to their behavioral characteristics. The granularity of this group evidence data is coarser than that of cognitive domain-level evidence. Therefore, the confidence level corresponding to the attached second confidence level label is usually lower than that of the first confidence level label. By introducing this coarsest-grained source of evidence, even in the extreme cold-start situation where cognitive domain-level evidence is unavailable, the system can still obtain alternative evidence with some basis, rather than completely losing evidence support.

[0066] Finally, the first or second alternative evidence is input into the subsequent logic as historical evidence data.

[0067] It should be added that the confidence level markers can be used by subsequent logic, in conjunction with the aforementioned evidence sparsity penalty mechanism, to apply stronger prudence to the corresponding configurations, ensuring that the use of alternative evidence is commensurate with its credibility; step-by-step queries can stop backtracking to coarser granularities once the first alternative evidence has been obtained. Therefore, even in cold-start scenarios, the system can still obtain alternative evidence with granularity as close as possible to the target patient to support routing decisions, avoiding routing failures or decision-making errors due to missing evidence, thus improving routing availability and decision reliability in cold-start scenarios.

[0068] Furthermore, the system also includes an early calibration module to quickly gather evidence and save on acquisition overhead before the calibration-aware routing module calculates utility values.

[0069] The first step is to obtain a preset minimum calibration configuration set before the calibration sensing routing module calculates the utility value.

[0070] The minimum calibration configuration set is a set of pre-selected, limited-quantity preset configurations that can be preset for different cognitive domains, thereby limiting the number of evaluations required in the calibration stage from the source.

[0071] The second step is to sort the multiple preset configurations in the minimum calibration configuration set in descending order based on the historical average utility value.

[0072] The historical average utility value is obtained by statistical analysis of historical evidence data accumulated in the configuration evidence library. When a preset configuration lacks historical data, it can be assigned a default sorting priority. Based on this, the preset configurations that have performed better on average in previous runs are evaluated first, thereby increasing the probability that the early stop mechanism will hit the optimal configuration in advance.

[0073] The third step is to send simulation evaluation commands to the configuration in the sorted order, and obtain the best and second-best utility values ​​generated by the current simulation evaluation in real time.

[0074] The system combines sequential evaluation with real-time data acquisition, allowing it to dynamically assess the leading position of the current optimal configuration relative to other configurations even before all preset configurations have been evaluated, providing real-time evidence for subsequent early termination decisions. Simulated evaluation can be conducted offline using simulated question-and-answer behavior data generated by the behavioral simulation modeling module, thus collecting evidence without interfering with the actual training process of the target patient.

[0075] It should be noted that the simulation evaluation does not rely on existing historical evidence data in the configuration evidence base. Instead, it runs each candidate collaborative configuration live on simulated answer behavior data, instantly collecting the generation quality and resource consumption of each configuration in this simulation evaluation. Based on this, it substitutes the data into the aforementioned utility value calculation logic to calculate the corresponding utility value on the spot, thereby obtaining the optimal and second-optimal utility values ​​generated in the current simulation evaluation. Correspondingly, when a preset configuration does not yet have a historical average utility value available for sorting, it is assigned a default sorting priority to participate in the sorting. This allows the early calibration to independently complete evidence collection and utility value calculation even in the initial state where there is no evidence in the configuration evidence base, and uses the collected evidence to initialize the configuration evidence base.

[0076] The fourth step is to calculate the difference between the optimal utility value and the second-best utility value. When the difference exceeds the preset early stop threshold, a test termination instruction is generated, and the configuration corresponding to the optimal utility value is updated to the candidate collaboration configuration.

[0077] It should be noted that the difference data reflects the leading position of the current optimal configuration relative to other evaluated configurations. The preset early stop threshold has a diversion effect on subsequent operations. The larger the value, the more significantly the current optimal configuration needs to lead before triggering early stop, and the higher the calibration cost. The smaller the value, the easier it is to trigger early stop, and the more evaluation cost is saved in the calibration phase. When the difference data exceeds this threshold, a test termination instruction is generated, ending the current round of simulation evaluation early. This ensures that resource savings are not only reflected in the subsequent routing and collaboration generation phases but also extended to the evidence collection calibration phase.

[0078] A configurable multi-agent negotiation module is used to acquire target collaboration configuration data, activate the corresponding agent role according to the target collaboration configuration data, perform multiple rounds of interactive processing on the profile parameters, generate initial game design data, and calculate the corresponding consensus score.

[0079] like Figure 2 The diagram shows a flowchart of the three-party negotiation generation process provided in this application embodiment. This module selectively activates corresponding intelligent agent roles according to the target collaboration configuration data. The activated roles then perform multiple rounds of interactive processing around the profile parameters, jointly producing initial game design data. This effectively translates the on / off combination state data determined in the routing phase into the activation and deactivation of the corresponding roles. It should be noted that the multi-round interactive processing can be implemented through three-party negotiation: the coding role implements the content, the quality review role scores the data, and the cognitive effectiveness review role checks the scientific validity of the training. The process iterates under a preset maximum number of feedback rounds, and if this constraint is exceeded, the current optimal result is taken. The consensus score is obtained by a weighted sum of the scores from the cognitive analysis dimension and the quality review dimension, for example, with the cognitive analysis dimension having a higher weight and the quality review dimension having a lower weight. This score is used to characterize the overall quality of the generated result and is sent back to the configuration evidence base as part of the audit log. Therefore, the system can ensure generation quality through multi-role collaboration under high-intensity configurations, and reduce resource consumption by reducing the number of roles and interaction rounds under low-intensity configurations.

[0080] The cognitive analysis dimension score is given by the cognitive analysis role based on the degree of fit between the generated results and the clinical training needs data according to a preset scoring scale. The quality review dimension score is given by the quality review role based on the content quality, playability and other dimensions according to a preset scoring scale. The two scores are weighted and summed according to preset weights to obtain the consensus score.

[0081] Furthermore, the configurable multi-agent negotiation module includes a requirements analysis submodule, a game design submodule, and a review and combination submodule:

[0082] First, the requirements analysis submodule is used to receive profile parameters and transform and output clinical training requirements data when the target collaboration configuration data indicates that it is enabled.

[0083] The requirements analysis submodule is responsible for clinical requirements analysis and cognitive effectiveness constraints. Based on the target training domain and cognitive ability scores, it outputs clinical training requirements data, ensuring the generated content is clinically relevant and cognitively effective. It should be noted that this submodule is only activated when the target collaboration configuration data is enabled; when disabled, this requirements analysis step is omitted to reduce the corresponding model calls and interaction overhead.

[0084] Next, the game design submodule is used to acquire clinical training needs data and generate candidate game draft data accordingly.

[0085] The game design submodule is responsible for designing game mechanics and playability, transforming clinical training requirements data into specific game mechanics, gameplay, and content frameworks. When the requirements analysis submodule is closed and does not output clinical training requirements data, the game design submodule can directly generate candidate game draft data based on the profile parameters to ensure that the generation process can still proceed normally under low-intensity configuration.

[0086] Then, the review combination submodule is used to receive candidate game draft data and perform multi-dimensional scoring calculations when the target collaboration configuration data indicates that it is enabled, and output the inspection results; when the inspection result is passed, the candidate game draft data is marked as the initial game design data; when the inspection result is rejected, a regeneration instruction is triggered based on the preset maximum feedback round threshold.

[0087] It should be noted that the review panel submodule assumes the role of quality reviewer, responsible for multi-dimensional quality constraints. It scores content based on factors such as content quality, playability, and relevance to training requirements, and outputs the verification results. When closed, the draft is used directly as the generated result without verification, thus eliminating review overhead. A preset maximum feedback round threshold constrains the iteration process: if the number of iteration rounds does not reach this threshold, the system can continue to trigger regeneration to strive for a better result; if the number of iteration rounds reaches this threshold, the system stops iterating and uses the currently obtained best draft as the generated result, thereby avoiding iteration non-convergence and continuous resource consumption due to repeated rejections.

[0088] Furthermore, the review portfolio submodule also includes an issue classification and routing unit, used to differentiate routing based on the nature of the issue when the review fails:

[0089] The first step is to extract the error classification identifier from the test result when the test result is "fail".

[0090] Among them, the error classification label reflects the nature of the problem corresponding to the failure, providing a basis for subsequent differentiated routing by category. This enables the system to identify whether the problem originates from the cognitive science level, the game design level, or the program code level, so that the corrective action is applied to the real problem link, rather than redoing the entire generated result indiscriminately.

[0091] The second step is to generate a concession execution instruction and send it to the internal nodes of the system to perform cognitive logic concession processing if the error classification identifier indicates a cognitive science error.

[0092] Among them, cognitive logic concession processing is used to make appropriate adjustments and compromises to the relevant requirements at the cognitive level when cognitive science constraints conflict with other constraints and are difficult to satisfy simultaneously, so as to avoid the iteration failing to converge due to overly strict cognitive constraints.

[0093] The third step is to generate simplified execution instructions and feed them back to the game design submodule to reduce the game complexity if the error classification indicator indicates a game design error.

[0094] The game complexity downgrading process is executed by the game design submodule. It reduces the complexity of game mechanics when the game design is too complex to implement or meet quality requirements, making the candidate game draft data more reasonable in terms of feasibility and playability. Simplification instructions are then fed back to the game design submodule, allowing corrective actions to be addressed at the point where the problem arose.

[0095] The fourth step is to generate a repair execution instruction if the error classification identifier indicates a program code error and send it to the preset encoding submodule to perform code reconstruction processing.

[0096] It should be noted that the coding submodule corresponds to the coding role responsible for the specific content implementation, while the code reconstruction process is used to repair or regenerate defective program code, ensuring that the game content can be correctly implemented and executed. Thus, three different types of problems—cognitive science, game design, and program code—are directed to their respective processing, allowing each feedback iteration to correct the real problem areas, improving the targeting and convergence efficiency of iterative corrections, reducing invalid duplicate generation, and controlling collaborative interaction overhead while ensuring generation quality. When a single test result involves multiple types of errors, they can be processed sequentially according to a preset priority or in parallel.

[0097] Furthermore, the system also includes a behavioral simulation modeling module, which communicates with the grouped database to provide a classification basis for matching the second alternative evidence in the aforementioned cold start rollback process.

[0098] The first step is to obtain the real patient long-term learning trajectory sequence of the external system, and call the nonlinear least squares fitting algorithm to numerically fit the long-term learning trajectory sequence to extract feature behavior data with convergence characteristics.

[0099] The long-term learning trajectory sequence of real patients reflects the changes in their performance over time during long-term training and can be derived from real training records accumulated by existing cognitive training platforms. The logic behind employing a nonlinear least squares fitting algorithm is that the learning trajectory of real patients typically exhibits a nonlinear changing pattern. This algorithm can determine curve parameters by minimizing the sum of squared deviations between the fitted curve and the real trajectory. Feature behavior data with convergence characteristics are used to characterize the behavioral patterns that cause the learning curve to stabilize as the training progresses. This fitting algorithm can also be implemented using other curve fitting or regression methods capable of parametrically modeling the learning trajectory.

[0100] The second step is to generate multi-class patient behavior prototype data based on the clustering of characteristic behavior data. The multi-class patient behavior prototype data includes stable parameters, fluctuating parameters, declining parameters and fatigue parameters.

[0101] Among them, the stable parameter reflects a generally stable behavioral pattern in learning performance, the fluctuating parameter reflects a behavioral pattern in learning performance that fluctuates, the declining parameter reflects a behavioral pattern in learning performance that gradually decreases with the training process, and the fatigue parameter reflects a behavioral pattern in learning performance that declines due to fatigue. By summarizing real patients into these representative behavioral prototypes, the patient group is structurally divided according to behavioral characteristics. This clustering can be implemented using conventional clustering algorithms, and the classification of behavioral prototypes can be added or adjusted as needed.

[0102] Specifically, clustering uses the curve parameters of the fitted characteristic behavior data, i.e., the learning curve, as the clustering feature. Patients with similar curve parameters are grouped into the same category, and the curve parameters corresponding to the cluster centers of each category are respectively used as stable parameters, fluctuating parameters, declining parameters, and fatigue parameters. That is, the behavior prototypes and their parameters are determined by the clustering results, rather than pre-setting fixed parameters.

[0103] The third step involves updating the grouping database with prototype data of various patient behaviors as the standard classification criterion to assist in the matching of second alternative evidence.

[0104] It is worth noting that after updating the grouping database with multi-category patient behavioral prototype data as the standard classification basis, the grouping database gains the ability to match and categorize patients based on behavioral prototypes. This allows the database to determine the behavioral prototype patient group to which the target patient belongs and provides corresponding group evidence data when there is neither historical evidence data of the target patient nor cognitive domain-level evidence. This ensures that the matching of second surrogate evidence is based on behavioral prototypes derived from real patient data, improving the rationality and reliability of surrogate evidence matching in cold-start scenarios. Furthermore, the multi-category patient behavioral prototype data can be periodically updated as real patients' long-term learning trajectory sequences accumulate.

[0105] The game execution and adaptive intervention module is used to receive initial game design data to generate cognitive training game data, collect game answering behavior data of target patients when executing cognitive training game data to dynamically calculate difficulty adjustment parameters and feed them back to the subsequent game generation process, and package the generated resource consumption data, activation component data and consensus score into audit logs and send them to the configuration evidence library for updating.

[0106] This module serves as the output and feedback link of the process. It receives the initial game design data and generates cognitive training game data that can be used by the target patients. During the target patients' execution, it collects game answering behavior data and dynamically calculates the difficulty adjustment parameters based on this data, feeding them back to the subsequent game generation process so that the training difficulty can be adaptively adjusted according to the patient's performance.

[0107] It should be noted that the resource consumption data records the resource consumption generated this time, and the activated component data records the actual combination of intelligent agent roles used this time. Both, along with the consensus score, are packaged into an audit log and sent to the configuration evidence repository. This audit log is the source of historical evidence data. By writing back the quality and resource consumption of each actual generation to the configuration evidence repository, the cumulative number of evidence entries for each candidate collaboration configuration continuously increases, and the evidence sparsity continuously decreases. This, in turn, affects the value of the evidence sparsity penalty weight and the calculation of the utility value in subsequent routing, so that the routing decision is continuously calibrated and tends to be more accurate as the operation progresses, ultimately forming a closed loop of dynamic balance between quality and resource consumption.

[0108] Furthermore, the game execution and adaptive intervention module includes a dynamic difficulty adjustment submodule, which adaptively adjusts the training difficulty based on the patient's real-time performance.

[0109] The first step is to analyze the target patients' accuracy tracking data in real time from the game's quiz behavior data.

[0110] Among them, the accuracy tracking data reflects the correctness of the target patients' answers during the cognitive training game and how it changes as the training progresses. It transforms the patients' actual answer performance into a state quantity that can be used for subsequent difficulty decisions, so that the difficulty adjustment can closely follow the patients' current performance.

[0111] The second step involves feeding the accuracy tracking data as state input data into a pre-defined dynamic detection reinforcement learning model, which contains algorithmic logic based on exploration and exploitation strategies.

[0112] It's important to explain that difficulty adjustment requires striking a balance between utilizing the currently known optimal difficulty (i.e., leveraging) and exploring new difficulties to discover a more suitable value (i.e., exploration). Therefore, by employing algorithmic logic based on exploration and utilization strategies, the system can both stably approximate the appropriate difficulty using existing experience and avoid getting stuck in locally unsuitable difficulty settings through appropriate exploration. Dynamic exploration reinforcement learning models can be implemented using a strategy of random exploration with a certain probability and selecting the current optimal action with the remaining probabilities, or by using other reinforcement learning strategies that combine exploration and utilization capabilities.

[0113] The third step is to use a dynamic detection reinforcement learning model to calculate and output the difficulty gain data for the next game round.

[0114] Among them, the difficulty gain data reflects the direction and magnitude of the difficulty adjustment that should be made in the next round relative to the current difficulty. When the accuracy rate is too high, it tends to output a gain that increases the difficulty, and when it is too low, it tends to output a gain that decreases the difficulty, thereby pushing the accuracy rate closer to the appropriate range.

[0115] Among them, the dynamic detection reinforcement learning model can be pre-trained offline using simulated answering behavior data generated by the behavior simulation modeling module, and continuously updated online based on real game answering behavior data in actual use.

[0116] The fourth step is to fuse the difficulty adjustment parameters with the difficulty gain data to generate updated difficulty adjustment parameters, so as to control the game design variables in the next cognitive training game data until the accuracy tracking data falls within the preset target accuracy range.

[0117] It should be noted that fusion calculation involves adding the adjustment amount calculated in this round to the current difficulty level to obtain the difficulty setting for the next round. This is used to control the game design variables in the next cognitive training game data. Game design variables can include adjustable elements that affect training difficulty, such as the difficulty of the questions, the time limit for answering, and the number of distractors. The above processing is continuously iterated during training. The preset target accuracy range is used to define the range of accuracy considered appropriate, acting as a diversion mechanism. When the accuracy falls into this range, it indicates that the current difficulty has become appropriate and the difficulty adjustment has stabilized. When the accuracy deviates from this range, the system continues to adjust the difficulty according to the direction of deviation to bring it back to convergence, thereby improving the adaptability and effectiveness of cognitive training.

[0118] It should be added that the updated difficulty adjustment parameter can be obtained by combining the current difficulty adjustment parameter and the difficulty gain data in a weighted summation manner. When the difficulty gain data is positive, the difficulty adjustment parameter is increased, and when it is negative, it is decreased. There is a preset mapping relationship between the difficulty adjustment parameter and each game design variable. For example, as the difficulty adjustment parameter increases, the difficulty level of the questions is increased, the answering time limit is shortened, and the number of distractors is increased, so that the change of the difficulty adjustment parameter is implemented as the specific value adjustment of each game design variable.

[0119] Considering that various types of data, such as profile parameters, historical evidence data, and game quiz behavior data, are generated at different times and with different frequencies, this system aligns various types of data with a unified time benchmark during the utility value calculation and audit write-back process: historical evidence data is uniformly classified into the corresponding time window according to its generation timestamp and participates in the aforementioned time-weighted calculation; game quiz behavior data is statistically tracked for accuracy rate according to a fixed time window; and resource consumption data, activation component data, and consensus scores generated in each generation are packaged as units of the same generation. This ensures that routing decisions, difficulty adjustments, and audit feedback are based on a consistent time correspondence, avoiding judgment bias caused by inconsistent data timing.

[0120] It should be added that the fault tolerance mapping relationship of cost penalty weights, the decay form of evidence sparsity penalty weights, state penalty coefficients, time decay constants, preset hysteresis thresholds, preset early stopping thresholds, maximum feedback rounds thresholds, and target accuracy ranges can all be calibrated offline based on the statistical distribution of labeled historical training data and combined with on-site expert experience, and can be updated periodically as individual historical training data accumulates; the minimum calibration configuration set, the value method of each confidence label, and the error classification identifier division of the problem classification route are all pre-configured. The above calibration and pre-training methods are only illustrative examples and do not constitute a limitation on the specific implementation.

[0121] Through the above description of the implementation methods, those skilled in the art can clearly understand that, for the sake of convenience and brevity, only the division of the above functional modules is used as an example. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the above functions can be divided into different functional modules to complete all or part of the functions described above.

[0122] In the embodiments provided in this application, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the embodiments described above are merely illustrative; for instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another device, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, or indirect coupling or communication connection between devices or units, and may be electrical, mechanical, or other forms.

[0123] The units described as separate components may or may not be physically separate. A component shown as a unit can be one or more physical units, located in one place or distributed in multiple different locations. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0124] Furthermore, the functional units in the various embodiments of this application 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. The integrated unit can be implemented in hardware or as a software functional unit.

[0125] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a readable storage medium. Based on this understanding, the technical solutions of the embodiments of this application, or the parts that contribute to the solution, or all or part of the technical solution, can be embodied in the form of a software product. This software product is stored in a storage medium and includes several instructions to cause a device (which may be a microcontroller, chip, etc.) or processor to execute all or part of the steps of the methods of the various embodiments of this application. 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.

[0126] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A multi-agent cognitive training game generation system for cognitive impairment, characterized in that, include: The profile parameter acquisition module is used to acquire profile parameters of the target patient, including the target training domain and cognitive ability score. The calibration perception routing module is used to receive the profile parameters, extract historical evidence data of each candidate collaboration configuration from the pre-built configuration evidence library, determine adaptive cost penalty weights based on the target training domain, calculate evidence sparsity penalty weights based on the cumulative number of evidence items in the historical evidence data, and calculate the utility value of each candidate collaboration configuration by combining the historical evidence data and the cost penalty weights. The utility value is processed according to the preset judgment logic to determine the target collaboration configuration data, wherein the candidate collaboration configuration is the combination of the on or off states of multiple intelligent agent roles in the system. A configurable multi-agent negotiation module is used to acquire the target collaboration configuration data, activate the corresponding agent role according to the target collaboration configuration data, perform multiple rounds of interactive processing on the profile parameters, generate initial game design data, and calculate the corresponding consensus score. The game execution and adaptive intervention module is used to receive the initial game design data to generate cognitive training game data, collect game answering behavior data of the target patient when executing the cognitive training game data to dynamically calculate the difficulty adjustment parameters and feed them back to the subsequent game generation process, and package the generated resource consumption data, activation component data and the consensus score into audit logs and send them to the configuration evidence library for updating.

2. The multi-agent cognitive training game generation system for cognitive impairment as described in claim 1, characterized in that: The calibration-aware routing module includes a utility calculation submodule, which is used for: Historical quality assessment data and historical resource consumption data of each candidate collaboration configuration are extracted from the historical evidence data. Establish a positive correlation between the fault tolerance of the target training domain and the cost penalty weight, so as to output the corresponding cost penalty weight; Establish a negative correlation between the cumulative number of evidence items and the evidence sparsity penalty weight, so as to output the corresponding evidence sparsity penalty weight; The utility value of each candidate collaboration configuration is calculated by subtracting the product of the historical resource consumption data and the cost penalty weight from the historical quality assessment data, and by subtracting the product of the set state penalty coefficient and the evidence sparsity penalty weight.

3. The multi-agent cognitive training game generation system for cognitive impairment as described in claim 1, characterized in that: The calibration-aware routing module includes a hysteresis determination submodule, which is used to execute the preset determination logic: The calculated utility values ​​are compared, and the candidate collaboration configuration corresponding to the maximum value is extracted as the first collaboration configuration data. Obtain the second collaboration configuration data used in the previous game generation round, and extract the reference utility value corresponding to the second collaboration configuration data; Calculate the difference between the utility value of the first collaborative configuration data and the reference utility value; When the difference data is greater than the preset hysteresis threshold, the first collaborative configuration data is confirmed as the target collaborative configuration data; When the difference data is not greater than the preset hysteresis threshold, the second collaboration configuration data is kept as the target collaboration configuration data.

4. The multi-agent cognitive training game generation system for cognitive impairment as described in claim 1, characterized in that: The system also includes an early calibration module, which is used for: Before the calibration-aware routing module calculates the utility value, a preset minimum calibration configuration set is obtained; The preset configurations in the minimum calibration configuration set are sorted in descending order based on the historical average utility value. The simulation evaluation commands are sent to the configuration in the order of sorting, and the best and second-best utility values ​​generated by the current simulation evaluation are obtained in real time. Calculate the difference between the optimal utility value and the suboptimal utility value. When the difference is greater than a preset early stop threshold, generate a test termination instruction and update the configuration corresponding to the optimal utility value to the candidate collaboration configuration.

5. The multi-agent cognitive training game generation system for cognitive impairment as described in claim 1, characterized in that: The configurable multi-agent negotiation module includes a requirements analysis submodule, a game design submodule, and a review and combination submodule, wherein: The requirements analysis submodule is used to receive the profile parameters and convert and output clinical training requirements data when the target collaboration configuration data indicates that it is enabled. The game design submodule is used to acquire the clinical training requirements data and generate candidate game draft data accordingly. The review panel submodule is used to receive the candidate game draft data and perform multi-dimensional scoring calculations when the target collaboration configuration data indicates that it is enabled, and output the verification results. When the test result is passed, the candidate game draft data is marked as the initial game design data; When the test result is rejection, a regeneration instruction is triggered based on a preset maximum feedback round threshold.

6. The multi-agent cognitive training game generation system for cognitive impairment as described in claim 5, characterized in that: The review combination submodule further includes a problem classification and routing unit, which is used for: When the test result is "fail", extract the error classification identifier from the test result; If the error classification identifier indicates a cognitive science error, a concession execution instruction is generated and sent to an internal system node to perform cognitive logic concession processing; If the error classification identifier indicates a game design error, a simplified execution instruction is generated and fed back to the game design submodule for game complexity downgrading. If the error classification identifier indicates a program code class error, a repair execution instruction is generated and sent to a preset encoding submodule to perform code reconstruction processing.

7. The multi-agent cognitive training game generation system for cognitive impairment as described in claim 1, characterized in that: The calibration-aware routing module further includes a cold start rollback submodule, which is used to perform the following processing when extracting the historical evidence data: The system checks whether there is historical evidence data of the target patient in the configured evidence base that completely matches the profile parameters. If not, it performs a step-by-step query according to priority. Query whether there is cognitive domain-level evidence in the configured evidence base that matches the target training domain. If it exists, extract it as the first alternative evidence and attach a first confidence label. If the cognitive domain-level evidence does not exist, the profile parameters are sent to a preset group database for matching, the corresponding behavioral prototype patient group data is obtained, the group evidence data is extracted as the second alternative evidence, and a second confidence level label is attached. The first or second alternative evidence is input into the subsequent logic as the historical evidence data.

8. The multi-agent cognitive training game generation system for cognitive impairment as described in claim 7, characterized in that: The system also includes a behavior simulation modeling module, which is communicatively connected to the grouped database and is used for: Obtain the real patient long-term learning trajectory sequence from the external system, and call the nonlinear least squares fitting algorithm to numerically fit the long-term learning trajectory sequence to extract feature behavior data with convergence characteristics. Based on the characteristic behavior data, multiple types of patient behavior prototype data are generated by clustering. The multiple types of patient behavior prototype data include stable parameters, fluctuating parameters, declining parameters, and fatigue parameters. The multi-type patient behavior prototype data are used as a standard classification criterion to update the grouping database to assist in the matching of the second alternative evidence.

9. The multi-agent cognitive training game generation system for cognitive impairment as described in claim 2, characterized in that: The utility calculation submodule further includes a time-weighting unit, which is used for: Obtain the generation timestamp corresponding to each piece of evidence in the historical quality assessment data, and calculate the difference between the generation timestamp and the current system time; Calculate the time decay coefficient based on the difference data, and construct a negative correlation function between the difference data and the time decay coefficient; The time decay coefficient is used to perform a weighted summation calculation on each of the historical quality assessment data to generate corrected quality assessment data, and the corrected quality assessment data is used to replace the original historical quality assessment data in the calculation of the utility value.

10. The multi-agent cognitive training game generation system for cognitive impairment as described in claim 1, characterized in that: The game execution and adaptive intervention module includes a dynamic difficulty adjustment submodule, which is used for: The accuracy tracking data of the target patient is analyzed in real time from the game quiz behavior data; The accuracy tracking data is used as state input data and fed into a preset dynamic detection reinforcement learning model, which contains algorithmic logic based on exploration and exploitation strategies. The dynamic detection reinforcement learning model calculates and outputs the difficulty gain data for the next game round. The difficulty adjustment parameters are fused with the difficulty gain data to generate updated difficulty adjustment parameters, which are used to control the game design variables in the next cognitive training game data until the accuracy tracking data falls within the preset target accuracy range.

Images