Education training scoring and optimization method based on multi-agent process collaborative training

Abstract

The application discloses an education training scoring and optimization method based on multi-agent process collaboration training, and relates to the technical fields of intelligent education and multi-agent collaboration. The method constructs a collaborative system composed of a training guidance agent, a special evaluation agent group, an adversarial feedback agent and a global coordination agent. Through subtask decomposition of the education training task, combined with multi-modal data such as code, voice and video, the training task is evaluated in a process and multi-dimensional manner. An adversarial feedback mechanism is introduced to correct the credibility of the special evaluation results and improve the scoring robustness. Through local continuity constraints and global consistency optimization, dynamic balance between subtask scoring and overall objectives is achieved. The method can realize real-time, traceability and iterative optimization of education training scoring, improve the objectivity, fairness and generalization ability of the evaluation results, and is suitable for various online education and vocational training scenarios.

Description

Technical Field

[0001] This invention relates to the fields of intelligent education and multi-agent collaboration technology, and in particular to an educational training scoring and optimization method based on multi-agent procedural collaborative training. Background Technology

[0002] With the popularization of online education and vocational training, students are generating increasingly rich data in training scenarios such as virtual experiments, programming training, and robot / drone operation, involving multimodal information such as code, voice, and video. How to objectively quantify and score students' operations, and on this basis, achieve dynamic optimization and personalized feedback, has become an important research problem in the field of educational technology.

[0003] Traditional practical training assessments primarily rely on manual evaluation or single-dimensional automated scoring, making it difficult to simultaneously consider multi-dimensional ability assessment, real-time feedback, and long-term learning trend prediction. This results in evaluation results lacking systematicity and sustainable optimization capabilities. Existing methods are mostly based on fixed templates or single-agent assessment systems, only capable of evaluating student operations from a single dimension or modality, such as focusing solely on code correctness or whether experimental operations conform to preset specifications, lacking a comprehensive measurement of multi-dimensional abilities. Furthermore, fixed scoring rules and the lack of dynamic adjustment mechanisms make it difficult to adapt to different students' learning habits and individual needs, easily leading to biased or distorted scoring results. Single-agent assessment methods also fail to achieve collaborative optimization between different evaluation dimensions and lack the ability to integrate multimodal information (such as code, voice, and video). Therefore, they have significant shortcomings in real-time feedback, iterative optimization, and long-term learning trend prediction, making it difficult to construct a complete closed-loop educational practical training system. Therefore, there is an urgent need for a closed-loop scoring and optimization method that can fully utilize multimodal data and combine multi-agent collaboration to achieve a real-time, accurate, traceable, and iteratively improveable intelligent practical training evaluation system. Summary of the Invention

[0004] To overcome the shortcomings of the prior art, this invention provides an educational training scoring and optimization method based on multi-agent procedural collaborative training, which evaluates the training process in a procedural and multi-dimensional manner.

[0005] To achieve the above objectives, the present invention adopts the following technical solution, including:

[0006] A method for scoring and optimizing educational training based on multi-agent procedural collaborative training. The method involves models including: training guidance agent, special evaluation agent group, adversarial feedback agent, and global coordination agent.

[0007] The training guidance agent breaks down educational training tasks into multiple sub-tasks. The training user executes each sub-task and obtains the execution results of each sub-task. The execution results of the sub-tasks are multimodal data.

[0008] Each specialized evaluation agent in the specialized evaluation agent group performs specialized scoring on the execution results of each subtask, resulting in specialized scores for each subtask; a subtask score is obtained based on the specialized scores of each subtask; and a comprehensive score for the training user is obtained based on the scores of each subtask.

[0009] Adversarial feedback agents are used to perform adversarial feedback on various subject scores. For each subject score, a subject feedback score is calculated. Based on the relationship between each subject score and the subject feedback score, the adversarial feedback loss is calculated.

[0010] The global coordinating agent is responsible for establishing a coordination mechanism among all subtask scores. On the one hand, it constrains the scores of adjacent subtasks based on their correlation and calculates the local evaluation loss at the subtask level. On the other hand, it constrains the consistency between the scores of each subtask and the accuracy of the comprehensive score and calculates the global coordination loss.

[0011] During training, the total model loss is calculated based on adversarial feedback loss, local evaluation loss, and global coordination loss, and the model parameters are optimized.

[0012] The trained model is used to perform process-based and comprehensive scoring on the educational training tasks of practical users, and behavioral trajectories are generated and stored.

[0013] Preferably, the special evaluation agent group includes a code quality evaluation agent, a function implementation evaluation agent, an operational efficiency evaluation agent, and a specification review evaluation agent;

[0014] Each specialized evaluation agent outputs its corresponding specialized score based on the multimodal coding features of the multimodal data:

[0015] ;

[0016] Among them, the multimodal data Including code text Video visual content and voice interaction content First, put the code text Video visual content and voice interaction content Input the corresponding text encoder, visual encoder, and audio encoder respectively, and perform encoding preprocessing to obtain the corresponding encoded features. The superscript z indicates a mode. Then, the encoded features are input into each specialized evaluation agent for specialized evaluation; The output score of the d-th special evaluation agent is the d-th special evaluation score. Let be the special evaluation function of the d-th special evaluation agent; d represents the d-th special evaluation agent, d=1,2,...,M; M is the total number of special evaluation agents.

[0017] Preferably, for the d-th specific score Corresponding special feedback score The expression is:

[0018] ;

[0019] in, H represents the adversarial function; H is the experience score, determined by the educational training task. The overall average score is calculated by averaging the average of the historical scores of the training users; during the training process, if the training users have no historical scores, the average historical scores are not included in the calculation of the experience score H; educational training tasks The overall average score is provided by the dataset annotations;

[0020] Scoring of the nth subtask ,in, The weight of the d-th special evaluation agent;

[0021] Overall rating of training users The educational training tasks are divided into N sub-tasks. Let n be the weight of the nth subtask, where n = 1, 2, ..., N.

[0022] Preferably, the adversarial feedback loss function is as follows:

[0023] ;

[0024] in, The output score of the d-th subject evaluation agent is the d-th subject score; d=1,2,...,M; M is the total number of subject evaluation agents; For the specific scoring Corresponding specific feedback scores; Let d be the adversarial feedback weight for the d-th special evaluation agent; It is a positive number; To counteract feedback loss.

[0025] The preferred local evaluation loss function is as follows:

[0026] ;

[0027] in, For localized assessment of loss; , The scores are for the nth and (n-1th)th subtasks, respectively. The educational training task is divided into N subtasks. For indicator functions, if and ,but =1, otherwise =0; where, For adjacent subtasks and The correlation score For correlation threshold, The threshold for the difference in scores between subtasks.

[0028] Preferably, the global coordination loss function is as follows:

[0029] ;

[0030] in, To mitigate overall losses, the educational training task was divided into N sub-tasks. The variance among the scores of all subtasks; The training users will be given a comprehensive score. Let n be the weight of the nth subtask. The score for the nth subtask, where n = 1, 2, ..., N; The actual ratings from training users are provided by the dataset annotations; , This is a hyperparameter.

[0031] Preferably, the training guides the intelligent agent to automatically generate and optimize sub-tasks by utilizing the knowledge and reasoning capabilities of the large model; the correlation between adjacent sub-tasks is calculated by the large model.

[0032] Preferably, the total loss function is:

[0033] ;

[0034] in, It is the set of all learnable parameters, i.e., model parameters; To adjust the parameters; Total loss; Local assessment of loss; The global coordination loss function; To counteract feedback loss, the educational training tasks are divided into N sub-tasks, n=1,2,...,N.

[0035] The present invention also provides an electronic device, characterized in that it includes a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it implements the educational training scoring and optimization method based on multi-agent procedural collaborative training.

[0036] The present invention also provides a computer program product, characterized in that it includes a computer program / instruction, which, when executed by a processor, implements the aforementioned educational training scoring and optimization method based on multi-agent procedural collaborative training.

[0037] The advantages of this invention are:

[0038] (1) Traditional scoring systems mostly rely on a single evaluation index, which makes it difficult to comprehensively measure students’ true abilities. In response to the problems of complex multimodal data, diverse evaluation dimensions and insufficient scoring stability in educational training scenarios, this invention constructs a collaborative system composed of training guidance agents, special evaluation agent groups, adversarial feedback agents and global coordination agents. Through the division of labor among multiple agents, a closed-loop collaboration of task decomposition, process tracking and global optimization is achieved, ensuring the multidimensionality and dynamic optimizability of the evaluation.

[0039] (2) This invention decomposes the educational training tasks into sub-tasks and combines multi-modal data such as code, voice and video to conduct process-oriented and multi-dimensional evaluation of the training tasks.

[0040] (3) Unlike existing methods that directly adopt scoring results, this invention designs an adversarial feedback agent and introduces an adversarial feedback mechanism. Based on the robust scoring enhanced by adversarial feedback, the evaluation model can continuously correct itself when faced with historical experience and current data, significantly improving the generalization of the scoring. Through this adversarial interaction process, the scoring system can effectively avoid biases caused by a single data source or accidental factors, making the final results more objective and credible, thus ensuring that the scoring system still has high reliability and promotional value in complex and ever-changing training environments.

[0041] (4) This invention introduces a global coordination agent optimization strategy to achieve a dynamic balance between multi-dimensional scoring and the global objective, avoiding the problem of over-optimization of a single objective and improving the comprehensiveness and scientific nature of the evaluation. During the optimization process, the global coordination agent automatically adjusts the weights of each specific evaluation item, ensuring that the overall objective function can simultaneously consider local accuracy and global consistency. This guarantees that the scoring results are both fair and comprehensive, and can truly reflect the overall level of the training process.

[0042] (5) The method of the present invention can realize the real-time, traceability and iterative optimization of educational training scores, improve the objectivity, fairness and generalization of evaluation results, and is applicable to various types of online education and vocational training scenarios. Attached Figure Description

[0043] Figure 1 This is a model architecture diagram of the present invention.

[0044] Figure 2 This is a schematic diagram of an embodiment of the present invention. Detailed Implementation

[0045] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0046] This invention proposes an educational training scoring and optimization method based on multi-agent procedural collaborative training. The method involves a model comprising: a training guidance agent, a group of specialized evaluation agents, an adversarial feedback agent, and a global coordination agent. These agents have clearly defined roles and collaborate to complete the entire process from task decomposition, process recording, specialized evaluation to global optimization. The model architecture and overall process are as follows: Figure 1 As shown.

[0047] The educational training scoring and optimization method based on multi-agent procedural collaborative training includes the following steps:

[0048] S1, Task Breakdown and Practical Training Guidance.

[0049] The training guidance agent breaks down the educational training task T (such as drone formation task, robot dance task, 3D robotic arm operation task) into a set of N sub-tasks:

[0050] ;

[0051] in, Let n represent the nth subtask, where n = 1, 2, ..., N.

[0052] Each subtask is assigned to an external execution module by the training guidance agent. The training user (student) completes each subtask in the execution module and transmits the execution results of each subtask to the subsequent special evaluation agent group.

[0053] The training guidance agent can not only automatically generate and optimize sub-tasks by utilizing the knowledge and reasoning capabilities of large models, but also allows training administrators (teachers) to flexibly customize sub-task planning according to teaching objectives.

[0054] The execution module is an interactive functional module that undertakes the sub-tasks assigned by the training guidance agent and supports training users (students) in completing practical training.

[0055] S2, for the execution results of each subtask of the training user, the special evaluation intelligent agent group conducts special evaluation to obtain the special score of each subtask; the subtask score is obtained based on the special score of each subtask; and the comprehensive score of the training user is obtained based on the scores of each subtask.

[0056] The special evaluation agent group includes multiple special evaluation agents, which evaluate the special scores of the sub-task execution results from different perspectives. In this embodiment, there are 4 special evaluation agents: (1) code quality agent: analyzes code structure and complexity; (2) function implementation agent: verifies whether the function is achieved; (3) running efficiency agent: evaluates the time / space efficiency of the algorithm; (4) specification review agent: checks naming conventions, comments and documentation.

[0057] Step S2 is detailed below:

[0058] S21, the output expressions of each special evaluation agent for the subtask execution results are as follows:

[0059] ;

[0060] The sub-task execution results submitted by the training users are multimodal data. Including code text Video visual content and voice interaction content , First, put the code text Video visual content and voice interaction content Input the corresponding text encoder, visual encoder, and audio encoder respectively, and perform encoding preprocessing to obtain the corresponding encoded features. The superscript z indicates a mode. Then, the encoded features are input into each specialized evaluation agent for specialized evaluation; Let d be the special evaluation function for the d-th special evaluation agent; The output score of the d-th special evaluation agent is the d-th special evaluation score; d represents the d-th special evaluation agent, d=1,2,...,M; M is the total number of special evaluation agents, and in this embodiment M=4.

[0061] In this invention, the text encoder and visual encoder of the CLIP model are used to process the code text. and video visual content Encoding is performed, and the audio encoder using the CLAP model is used to process the voice interaction content. Encoding is performed. CLIP and CLAP models are external perception modules used to assist agents in understanding multimodal data in the external environment.

[0062] The CLIP model is an image-text contrastive pre-trained model. In the CLIP model, the visual encoder (image encoder) and text encoder are independent, used to encode images and text respectively, yielding image-encoded features and text-encoded features. The CLAP model is an audio-text contrastive pre-trained model. In the CLAP model, the audio encoder and text encoder are independent, used to encode audio and text respectively, yielding audio-encoded features and text-encoded features.

[0063] S22, Based on the scores of each specific item in the nth subtask, obtain the score of the nth subtask. ,in, The weights for the d-th special evaluation agent are learned adaptively through model adaptation.

[0064] S23. Based on the scores of each sub-task of the training user, obtain the comprehensive score of the training user. ,in, The weight of the nth subtask can be preset by the training administrator (teacher) or learned adaptively through the model.

[0065] S3, the adversarial feedback agent performs adversarial feedback, and provides a specific feedback score for each specific score; based on each specific score and the corresponding specific feedback score, the adversarial feedback loss is calculated.

[0066] During training, a group of specialized evaluation agents provides specialized scores for each subtask. Then, the model does not directly accept these scores; an adversarial process is added to evaluate these subject scores. The credibility level in the current task scenario is recorded as a specific feedback score. The special feedback score A training phase is introduced to enhance the robustness and personalization of the scoring. Specific feedback scoring. The expression is:

[0067] ;

[0068] in, This indicates that the adversarial feedback agent is responsible for specific scoring. Specific feedback scores; H represents the adversarial function; H is the experience score, determined by the educational training task. The overall average score is calculated by averaging the average of the historical scores of the training users. During the training process, if a training user has no historical scores, the average historical scores are not included in the calculation of the experience score H; and the educational training task... The overall average score is provided by the dataset annotations.

[0069] Based on the specific feedback score output by the adversarial feedback agent, this invention designs an adversarial feedback loss function. Used to measure the scores of each subject. Corresponding special feedback scores The difference between them is expressed as:

[0070] ;

[0071] in, Let be the adversarial feedback weights of the d-th specialized evaluation agent, and be the learnable parameters during the training phase; It is a very small positive constant to prevent excessively small values ​​from appearing during the adversarial process; this loss function This is used to constrain the accuracy of each special evaluation agent on its own scoring task. The optimization objective is: when the difference is large, the special evaluation agent optimizes its scoring model to make it closer to the refutation of reasonableness; when the difference is small, it indicates that the special evaluation and the adversarial feedback have reached an agreement, indicating that the evaluation results are more robust.

[0072] S4, the global coordinating agent, is responsible for establishing a coordination mechanism among all sub-task scores to achieve global optimization. Its core lies in balancing the rationality of dividing the educational training task T into N sub-tasks and improving the accuracy of the scores for each of the N sub-tasks. Its core lies in the design of a two-level loss function.

[0073] (1) Local evaluation loss function at the sub-task level To ensure the continuity and rationality of process-based scoring among interconnected sub-tasks, this invention introduces a scoring constraint mechanism based on the prediction of sub-task correlation using a large model during the local specialized evaluation stage. Adjacent sub-tasks are obtained through calculation by the large model or annotation by the training administrator (teacher). and Relevance score When it is above the correlation threshold The two adjacent subtasks are considered to be highly correlated in terms of semantic goals or functional implementation, and their constraint form is expressed as:

[0074] ;

[0075] in, This indicates an indicator function that returns 1 when the condition is true and 0 otherwise. For correlation threshold, The threshold for the difference in scores between subtasks.

[0076] When the high correlation condition is met ( And the subtask score change exceeds the threshold. ( When the current score is constrained, the adjacent subtasks are made to... and The scoring is constrained, and no abnormal jumps occur.

[0077] Accordingly, the local evaluation loss function for:

[0078] ;

[0079] in, , These are the scores for the nth and n-1th subtasks, respectively. This local evaluation loss effectively suppresses fluctuations caused by task decomposition or evaluation bias by limiting the unreasonable increase in scores of highly correlated subtasks, thereby improving the stability and fairness of process-based scoring.

[0080] (2) Global Coordination Loss Function This is used to measure the consistency between the scores of each subtask and the accuracy of the overall score:

[0081] ;

[0082] Among them, educational training tasks Divided into N subtasks, Variance among scores for all subtasks is used to ensure consistency of evaluation results; The training users will be given a comprehensive score. The weight of the nth subtask can be preset by the training administrator (teacher) or learned adaptively through the model; The actual ratings from training users are provided by the dataset annotations; , These are hyperparameters used to adjust the balance.

[0083] The global coordination agent introduces subtask correlation constraints and a global consistency loss function to limit the score changes of highly correlated subtasks and optimize the deviation between the fused score and the preset target score, thereby improving the stability and fairness of the overall training score.

[0084] S5 calculates the total loss during training and optimizes the model parameters.

[0085] The objective of the entire multi-agent cooperative optimization is expressed as:

[0086] ;

[0087] in, It is the set of all learnable parameters (model parameters) that need to be trained / optimized; To adjust the parameters; Total loss; To counteract feedback loss and prevent distortion of scores for each specific item; To locally assess the loss, it is used to constrain cases of anomalous jumps in scores at the subtask level; This is a globally coordinated loss function used to ensure the accuracy of the scoring task from a global perspective.

[0088] S6 utilizes the trained model to perform process-based and comprehensive scoring of the educational training of practical users, recording and storing the behavioral trajectories.

[0089] The system automatically saves the completion process of subtasks and the results of specific evaluations, forming a behavior trajectory. This allows the system to return the rating of each subtask for the training user, as well as the training process data. Stored as tuples in a long-term archive, these records are used for subsequent learning status monitoring and model iteration, serving as a long-term record for managing students' status during practical training. :

[0090] ;

[0091] in, This represents the nth subtask in the training process. This represents the d-th specific evaluation agent pair of sub-tasks. The scores, where n and d both take values ​​in the range of positive integers. It can be adjusted according to different training tasks.

[0092] Example 1

[0093] like Figure 2As shown, this embodiment targets the UAV autonomous cruise and obstacle avoidance programming training scenario. The overall educational training task T is broken down into five sub-tasks: UAV takeoff and basic control sub-task t1, waypoint path planning algorithm implementation sub-task t2, dynamic obstacle detection sub-task t3, autonomous obstacle avoidance and real-time trajectory adjustment sub-task t4, and safe landing and system integrity verification sub-task t5. Among them, the waypoint path planning algorithm implementation sub-task t2 and the dynamic obstacle detection sub-task t3 are highly correlated at the functional implementation level and are therefore identified as highly correlated sub-tasks. Using the educational training scoring and optimization method based on multi-agent procedural collaborative training of the present invention, the educational training task is scored procedurally and comprehensively. The scoring results are shown in Table 1 below. In this embodiment, both the procedural score and the comprehensive score are within a reasonable range, and the difference between the predicted score and the actual labeled score for each sub-task does not exceed 1 point. Moreover, the scores of the two highly correlated sub-tasks, t2 and t3, are constrained by the method of the present invention.

[0094] Table 1. Results of Process-Based Scoring and Overall Scoring

[0095]

[0096] The above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for scoring and optimizing educational training based on multi-agent procedural collaborative training, characterized in that, The methods involve models including: training guidance agents, special evaluation agent groups, adversarial feedback agents, and global coordination agents; The training guidance agent breaks down educational training tasks into multiple sub-tasks. The training user executes each sub-task and obtains the execution results of each sub-task. The execution results of the sub-tasks are multimodal data. Each specialized evaluation agent in the specialized evaluation agent group performs specialized scoring on the execution results of each subtask, resulting in specialized scores for each subtask; a subtask score is obtained based on the specialized scores of each subtask; and a comprehensive score for the training user is obtained based on the scores of each subtask. Adversarial feedback is applied to each subject score using an adversarial feedback agent. For each subject score, a subject feedback score is calculated. Based on each subject score and the subject feedback score, the adversarial feedback loss is calculated. The global coordinating agent is responsible for establishing a coordination mechanism among all subtask scores. On the one hand, it constrains the scores of adjacent subtasks based on their correlation and calculates the local evaluation loss at the subtask level. On the other hand, it constrains the consistency between the scores of each subtask and the accuracy of the comprehensive score and calculates the global coordination loss. During training, the total model loss is calculated based on adversarial feedback loss, local evaluation loss, and global coordination loss, and the model parameters are optimized. The trained model is used to perform process-based and comprehensive scoring on the educational training tasks of the training users, and the behavioral trajectory is generated and stored. The global coordination loss function is shown below: in, To mitigate overall losses, the educational training task was divided into N sub-tasks. The variance among the scores of all subtasks; The training users will be given a comprehensive score. Let n be the weight of the nth subtask. The score for the nth subtask, where n = 1, 2, ..., N; The actual ratings from training users are provided by the dataset annotations; , This is a hyperparameter.

2. The educational training scoring and optimization method based on multi-agent procedural collaborative training as described in claim 1, characterized in that, The specialized assessment agent group includes code quality assessment agent, function implementation assessment agent, operational efficiency assessment agent, and specification review assessment agent; Each specialized evaluation agent outputs its corresponding specialized score based on the multimodal coding features of the multimodal data: Among them, the multimodal data Including code text Video visual content and voice interaction content First, put the code text Video visual content and voice interaction content Input the corresponding text encoder, visual encoder, and audio encoder respectively, and perform encoding preprocessing to obtain the corresponding encoded features. The superscript z indicates a mode. Then, the encoded features are input into each specialized evaluation agent for specialized evaluation; The output score of the d-th special evaluation agent is the d-th special evaluation score. Let be the special evaluation function of the d-th special evaluation agent; d represents the d-th special evaluation agent, d=1,2,...,M; M is the total number of special evaluation agents.

3. The educational training scoring and optimization method based on multi-agent procedural collaborative training as described in claim 2, characterized in that, For the d-th specific score Corresponding special feedback score The expression is: in, H represents the adversarial function; H is the experience score, determined by the educational training task. The overall average score is calculated by averaging the average of the historical scores of the training users; during the training process, if the training users have no historical scores, the average historical scores are not included in the calculation of the experience score H; educational training tasks The overall average score is provided by the dataset annotations; Scoring of the nth subtask ,in, The weight of the d-th special evaluation agent; Overall rating of training users The educational training tasks are divided into N sub-tasks. Let n be the weight of the nth subtask, where n = 1, 2, ..., N.

4. The educational training scoring and optimization method based on multi-agent procedural collaborative training according to claim 1, characterized in that, The adversarial feedback loss function is as follows: in, The output score of the d-th subject evaluation agent is the d-th subject score; d=1,2,...,M; M is the total number of subject evaluation agents; For the specific scoring Corresponding specific feedback scores; Let d be the adversarial feedback weight for the d-th special evaluation agent; It is a positive number; To counteract feedback loss.

5. The educational training scoring and optimization method based on multi-agent procedural collaborative training according to claim 1, characterized in that, The local evaluation loss function is shown below: in, For localized assessment of loss; , The scores are for the nth and (n-1th)th subtasks, respectively. The educational training task is divided into N subtasks. For indicator functions, if and ,but =1, otherwise =0; where, For adjacent subtasks and The correlation score For correlation threshold, The threshold for the difference in scores between subtasks.

6. The educational training scoring and optimization method based on multi-agent procedural collaborative training according to claim 1, characterized in that, The training guides the intelligent agent to automatically generate and optimize sub-tasks by utilizing the knowledge and reasoning capabilities of the large model; the correlation between adjacent sub-tasks is calculated by the large model.

7. The educational training scoring and optimization method based on multi-agent procedural collaborative training according to claim 1, characterized in that, The total loss function is: in, It is the set of all learnable parameters, i.e., model parameters; To adjust the parameters; Total loss; For localized assessment of loss; Losses due to overall coordination; To counteract feedback loss, the educational training tasks are divided into N sub-tasks, n=1,2,...,N.

8. An electronic device, characterized in that, It includes a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the educational training scoring and optimization method based on multi-agent procedural collaborative training as described in any one of claims 1 to 7.

9. A computer program product, characterized in that, It includes a computer program / instruction that, when executed by a processor, implements the educational training scoring and optimization method based on multi-agent procedural collaborative training as described in any one of claims 1 to 7.

Images