Experiment teaching effect evaluation method based on reinforcement learning

Abstract

The invention relates to the technical field of teaching evaluation, in particular to an experimental teaching effect evaluation method based on reinforcement learning, and the method comprises the steps: obtaining multi-dimensional impact factors, such as a teaching environment, student characteristics and experimental characteristics, which affect the experimental teaching effect, carrying out the clustering and grouping of experimental courses, and forming course groups with similar characteristics; determining an initial weight of an evaluation parameter in combination with an analytic hierarchy process, performing preliminary evaluation on the group, and dynamically screening a target evaluation parameter suitable for the group according to an evaluation result; constructing a membership function of the target parameters, and quantifying the membership degree of each parameter in the evaluation level; and finally fusing the parameter weight and the membership degree, introducing a reinforcement learning agent to carry out dynamic optimization evaluation, automatically adjusting an evaluation model through a strategy learning and reward feedback mechanism, and outputting a final evaluation result of the teaching effect. According to the method, intelligentization and adaptive optimization of the evaluation process are realized, and the accuracy, objectivity and practicability of experimental teaching evaluation are improved.

Description

Technical Field

[0001] This invention relates to the field of teaching assessment technology, and in particular to a method for evaluating the effectiveness of experimental teaching based on reinforcement learning. Background Technology

[0002] With the rapid development of educational informatization and intelligentization, experimental teaching, as a crucial component of higher and vocational education for cultivating students' practical and innovative abilities, has received increasing attention for its scientific evaluation. Traditional methods for evaluating the effectiveness of experimental teaching mainly rely on static evaluation methods such as final exam scores, lab report grades, or subjective teacher assessments. These methods fail to comprehensively reflect students' multi-dimensional performance, including operational skills, logical thinking, teamwork, and knowledge application during the experiment. Furthermore, existing evaluation models generally neglect the combined impact of differences in the teaching environment (such as equipment configuration and safety conditions), individual student characteristics (such as cognitive level and learning style), and the inherent characteristics of the experimental project itself (such as complexity and risk level) on teaching effectiveness, resulting in evaluation results that lack specificity and accuracy.

[0003] In recent years, some studies have attempted to introduce the Analytic Hierarchy Process (AHP) and fuzzy comprehensive evaluation models to quantify experimental teaching. While these methods have improved the objectivity of the evaluation to some extent, they still suffer from problems such as highly subjective weight setting, unchanging evaluation parameters, and inability to adapt to dynamic changes in teaching. Furthermore, these methods typically treat each experimental course in isolation, lacking the ability to explore and utilize the regular characteristics of course groups under similar teaching situations, thus limiting the generalization ability and optimization potential of the evaluation models.

[0004] The above content is only used to help understand the technical solution of the present invention and does not represent an admission that the above content is prior art. Summary of the Invention

[0005] The main objective of this invention is to provide a method for evaluating the effectiveness of experimental teaching based on reinforcement learning. This method aims to address the technical problems of existing experimental teaching effectiveness evaluation methods, which are unable to comprehensively consider multiple influencing factors such as teaching environment, student characteristics, and experimental characteristics. Furthermore, the evaluation process is static, the weight setting is subjective, and there is a lack of dynamic optimization mechanisms, resulting in insufficient accuracy and adaptability of the evaluation results.

[0006] To achieve the above objectives, this invention provides a method for evaluating the effectiveness of experimental teaching based on reinforcement learning, the method comprising: A number of experimental courses are set up in the teaching system. Multiple influencing factors that affect the teaching effect of the experimental courses are obtained. Based on the influencing factors, the experimental courses are classified to obtain multiple course groups consisting of at least one experimental course. The influencing factors include teaching environment parameters, student characteristic parameters, and experimental characteristic parameters. Determine the preset evaluation parameters for assessing teaching effectiveness, and assign different parameter weight values ​​to the evaluation parameters according to the analytic hierarchy process. For the course group, a preliminary evaluation of the teaching effectiveness is conducted based on the parameter weight values, resulting in a group evaluation result composed of individual evaluation results corresponding to multiple experimental courses. Based on the group evaluation results, target evaluation parameters for evaluating the teaching effectiveness of the corresponding course groups are selected from the evaluation parameters; For each experimental course, a membership function corresponding to the target evaluation parameter is constructed, and the target evaluation parameter value corresponding to the experimental course is collected. Based on the membership function and the target evaluation parameter value, the membership degree of the target evaluation parameter value within the preset evaluation level is calculated. Based on the parameter weights and membership degrees, the teaching effectiveness of the experimental course is dynamically optimized and evaluated by a reinforcement learning agent to determine the final teaching effect of the experimental course.

[0007] Optionally, the step of classifying the several experimental courses according to the influencing factor to obtain multiple course groups consisting of at least one experimental course specifically includes: Based on the aforementioned impact factors, determine the impact factor values ​​corresponding to each of the experimental courses; From the aforementioned experimental courses, select designated experimental courses whose differences between the values ​​of each influencing factor are within a preset range to form a course group, thereby classifying the aforementioned experimental courses. The preset range is dynamically adjusted according to the reinforcement learning training stage. A more lenient threshold is used in the initial training stage, and a more stringent threshold is used in the convergence stage.

[0008] Optionally, before selecting a specific experimental course from the plurality of experimental courses, the method further includes: Based on the historical performance data of the aforementioned experimental courses, a corresponding learning behavior change curve is generated, and based on the learning behavior change curve, the teaching stage of the current experimental course is estimated through the reinforcement learning state recognition module. The range of differences between the influencing factor values ​​is dynamically adjusted according to the teaching stage; wherein, the range of differences is positively correlated with the students' adaptability at the teaching stage: the lower the adaptability, the wider the range of differences.

[0009] Optionally, the step of selecting target evaluation parameters from the evaluation parameters to evaluate the teaching effectiveness of the course group based on the group evaluation results specifically includes: The fluctuation value between individual evaluation results corresponding to any two experimental courses in the group evaluation results is determined by a reinforcement learning policy network, and the maximum fluctuation value is selected from the fluctuation values. If the maximum fluctuation value is within a preset range, the evaluation parameter is used as the target evaluation parameter; Otherwise, the reinforcement learning action selection module is invoked to combine the evaluation parameters to generate an evaluation parameter set. Based on the evaluation parameter set, the evaluation strategy is optimized through multiple rounds of iterative training until the fluctuation values ​​corresponding to the parameter set all meet the preset range. Then, the evaluation parameter set is used as the target evaluation parameter.

[0010] Optionally, after iterative evaluation via reinforcement learning, the method further includes: When multiple evaluation parameter sets have fluctuation values ​​that meet the preset range, calculate the cumulative reward value for each parameter set; Based on the cumulative reward value, the set of evaluation parameters with the highest reward value is selected as the target evaluation parameter; wherein, the cumulative reward value is calculated by the reinforcement learning agent based on the improvement of teaching effect and evaluation stability.

[0011] Optionally, constructing the membership function corresponding to the target evaluation parameters specifically includes: The evaluation levels corresponding to the target evaluation parameters are divided; wherein each evaluation level is mapped to a state space interval in reinforcement learning. For each target evaluation parameter, the Q-learning algorithm is used to optimize the level membership degree of the parameter in each level interval. The level membership degree is then re-standardized based on reward feedback to obtain the membership degree function corresponding to the target evaluation parameter.

[0012] Optionally, determining the final teaching effect corresponding to the experimental course specifically includes: An evaluation vector is generated based on the membership degree of the target evaluation parameter value within each evaluation level; The evaluation vectors of all target evaluation parameters are concatenated into a state feature matrix, which is then input into the reinforcement learning evaluation model. The reinforcement learning evaluation model performs fuzzy transformation and reward update through its policy network, and outputs a confidence vector for teaching effectiveness. Based on the optimal action selected by the reinforcement learning agent in the teaching effect confidence vector, the final teaching effect corresponding to the experimental course is determined.

[0013] Optionally, determining the final teaching effect of the experimental course based on the optimal action selected by the reinforcement learning agent in the teaching effect confidence vector specifically includes: The evaluation level with the highest probability value selected by the reinforcement learning agent is selected from the confidence vector of the teaching effect. If the selection probability value is lower than a preset threshold, the exploration mechanism is triggered to regenerate the evaluation parameter set and return to the iterative evaluation step; Otherwise, the assessment level will be taken as the final teaching effect, and the assessment results will be stored in the experience playback pool to optimize subsequent strategies.

[0014] Furthermore, to achieve the above objectives, the present invention also provides an experimental teaching effectiveness evaluation device based on reinforcement learning. The device includes: a memory, a processor, and an experimental teaching effectiveness evaluation program based on reinforcement learning stored in the memory and executable on the processor. The experimental teaching effectiveness evaluation program based on reinforcement learning is configured to implement the steps of the experimental teaching effectiveness evaluation method based on reinforcement learning as described above.

[0015] In addition, to achieve the above objectives, the present invention also provides a medium storing a reinforcement learning-based experimental teaching effectiveness evaluation program, wherein when the reinforcement learning-based experimental teaching effectiveness evaluation program is executed by a processor, it implements the steps of the reinforcement learning-based experimental teaching effectiveness evaluation method as described above.

[0016] This invention provides a reinforcement learning-based method for evaluating the effectiveness of experimental teaching. The method scientifically classifies and groups experimental courses by integrating teaching environment parameters, student characteristic parameters, and experimental characteristic parameters. It uses the analytic hierarchy process (AHP) to determine the weights of evaluation parameters and introduces a membership function to achieve fuzzy quantification of evaluation indicators. Furthermore, it utilizes a reinforcement learning agent to dynamically optimize and evaluate teaching effectiveness. This method not only overcomes the problems of strong subjectivity and static rigidity in traditional evaluation methods but also achieves adaptability and intelligence in the evaluation process. It can accurately reflect the teaching characteristics of different course groups, improve the accuracy and fairness of evaluation results, and support the dynamic adjustment of teaching strategies through a continuous learning mechanism, effectively enhancing the scientific, intelligent, and practical nature of experimental teaching effectiveness evaluation. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating an embodiment of the experimental teaching effectiveness evaluation method based on reinforcement learning according to the present invention.

[0018] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0019] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0020] Reference Figure 1, Figure 1 This is a flowchart illustrating an embodiment of the experimental teaching effectiveness evaluation method based on reinforcement learning according to the present invention.

[0021] In one embodiment, the method for evaluating the effectiveness of reinforcement learning-based experimental teaching includes: Step S100: Determine several experimental courses set up in the teaching system, obtain multiple influencing factors affecting the teaching effectiveness of the experimental courses, and classify the experimental courses according to the influencing factors to obtain multiple course groups consisting of at least one experimental course. The influencing factors include teaching environment parameters, student characteristic parameters, and experimental characteristic parameters.

[0022] The teaching system can be an information platform or a physical-digital integrated environment that supports the implementation and management of experimental courses. It can provide the basic operating environment for experimental course setup, data collection, and evaluation. Several experimental courses can be specific experimental teaching units configured within the teaching system to cultivate students' practical abilities, serving as the objects of evaluation methods and data sources. Influencing factors can be a set of multidimensional variables affecting the teaching effectiveness of experimental courses, serving as the basis for course classification and group modeling. Furthermore, influencing factors can include, but are not limited to, one or more of the following: teaching environment parameters, student characteristic parameters, and experimental characteristic parameters. Teaching environment parameters can be variables describing the physical and technical support conditions for experimental teaching, reflecting the potential impact of external environments such as equipment configuration and safety conditions on teaching effectiveness. Student characteristic parameters can be variables characterizing individual differences among students participating in experiments, reflecting the impact of cognitive level, learning style, etc., on students' experimental performance. Experimental characteristic parameters can be variables describing the inherent attributes of the experimental project itself, reflecting the requirements of the teaching process on intrinsic characteristics such as experimental complexity and hazard level. A course group can be a collection of one or more experimental courses that share similarities in teaching environment, student characteristics, and experimental features. It can be used to capture common patterns in similar teaching contexts and provide a unified and adaptable assessment framework for courses within the group. In an exemplary embodiment, a course group may include, but is not limited to, one or more of the following: high-risk, high-complexity experimental groups, basic verification experimental groups, and interdisciplinary integrated design experimental groups.

[0023] Identifying several experimental courses set up in the teaching system can be achieved by retrieving the configured list of experimental courses from the teaching system's course database. Furthermore, this operation can be implemented by directly accessing the course metadata within the teaching management system, thereby achieving the technical effect of clearly defining the scope of the evaluation objects. Obtaining multiple influencing factors affecting the teaching effectiveness of experimental courses can be achieved by extracting teaching environment parameters, student characteristic parameters, and experimental characteristic parameters from the teaching management system, student archives, and experimental project metadata. Furthermore, this operation can be completed through a structured data interface or log parsing module, thereby achieving the technical effect of providing multi-dimensional feature inputs for course classification. Classifying several experimental courses according to the influencing factors yields multiple course groups, which can be grouped according to the similarity of influencing factors using clustering algorithms or rule engines. In one specific embodiment, this operation can be achieved by using the K-means clustering algorithm to perform unsupervised grouping of standardized multi-dimensional influencing factors; for example, it can also be achieved by performing supervised classification of influencing factor combinations based on decision tree rules, thereby achieving the technical effect of forming course groups with consistent teaching contexts and improving the generalization ability of the evaluation model.

[0024] Step S200: Determine the preset evaluation parameters for assessing teaching effectiveness, and assign different parameter weight values ​​to the evaluation parameters according to the analytic hierarchy process.

[0025] The evaluation parameters can be a pre-defined set of multi-dimensional indicators used to measure the effectiveness of experimental teaching, which can form the basic indicator system for preliminary evaluation. The parameter weights can be relative importance values ​​assigned to each evaluation parameter through the analytic hierarchy process, which can be used to reflect the priority of different evaluation dimensions in the preliminary evaluation.

[0026] Determining the preset evaluation parameters for assessing teaching effectiveness can be achieved by loading a predefined set of multi-dimensional evaluation parameters from an evaluation indicator library. Furthermore, this operation can be implemented by calling standardized teaching evaluation indicator templates, thereby achieving the technical effect of establishing a structured evaluation dimension framework. Different parameter weights are assigned to the evaluation parameters according to the analytic hierarchy process (AHP), and a judgment matrix can be constructed through expert scoring, calculating eigenvectors and normalizing them to obtain the weight values. In a specific embodiment, this operation can be implemented through an interactive expert scoring interface combined with a consistency verification mechanism, thereby achieving the technical effect of reducing the bias of completely subjective weighting while retaining a certain objective structure.

[0027] Step S300: For the course group, conduct a preliminary evaluation of the teaching effectiveness based on the parameter weight values ​​to obtain a group evaluation result composed of the individual evaluation results corresponding to multiple experimental courses.

[0028] A single evaluation result can be a preliminary evaluation score calculated based on parameter weights for a single experimental course, and can be used as the basic unit to constitute the group evaluation result. A group evaluation result can be a set of results obtained after preliminary evaluation of all experimental courses within a course group, and can be used as a basis for selecting target evaluation parameters.

[0029] For course groups, a preliminary evaluation of teaching effectiveness is conducted based on parameter weight values. This involves weighting and aggregating the evaluation parameter values ​​of each experimental course within the group to obtain a preliminary score. Furthermore, this operation can be implemented using vectorized weighted summation, achieving the technical effect of generating benchmark evaluation results that can be used for parameter filtering. The resulting group evaluation results, composed of individual evaluation results for multiple experimental courses, allow for the aggregation of preliminary evaluation scores for all experimental courses within the same course group. For example, this operation can be implemented through database aggregation queries or in-memory array traversal, thereby achieving the technical effect of forming a group-level evaluation dataset.

[0030] Step S400: Based on the group evaluation results, select the target evaluation parameters from the evaluation parameters to evaluate the teaching effectiveness of the corresponding course groups.

[0031] The target evaluation parameters can be key evaluation indicators selected from the original evaluation parameters that have a significant differentiating ability for the teaching effectiveness of a specific course group. They can be used to focus on the evaluation dimensions most relevant to the current course group, thereby improving the relevance and efficiency of the evaluation. In an exemplary embodiment, the target evaluation parameters may include, but are not limited to, one or more of the following: operational standardization indicators, knowledge transfer ability indicators, and team collaboration effectiveness indicators.

[0032] By selecting target evaluation parameters from the evaluation parameters based on the group evaluation results, the discrimination or variance contribution of each evaluation parameter within the group can be analyzed, retaining highly sensitive parameters. In one specific embodiment, this operation can be achieved by calculating the correlation coefficient between each evaluation parameter and the group average effect, retaining parameters whose absolute values ​​are higher than a threshold; furthermore, principal component analysis (PCA) can also be used to extract the top k parameters with the strongest explanatory power for the group evaluation results, thereby achieving the technical effect of focusing on key evaluation dimensions and reducing the complexity of subsequent fuzzy quantization and reinforcement learning.

[0033] Step S500: For each experimental course, construct the membership function corresponding to the target evaluation parameter, collect the target evaluation parameter value corresponding to the experimental course, and calculate the membership degree of the target evaluation parameter value within the preset evaluation level based on the membership function and the target evaluation parameter value.

[0034] The membership function can be a mathematical function used to map the target evaluation parameter value to a preset evaluation level fuzzy set. It can be used to achieve fuzzy quantification of imprecise and nonlinear teaching performance data, enhancing the evaluation's tolerance to uncertainty. In a specific embodiment, the membership function can be one or more of the following, including but not limited to triangular membership functions, trapezoidal membership functions, and Gaussian membership functions. The target evaluation parameter value can be the actual observed value of the target evaluation parameter collected in a specific experimental course, which can be used as the input to the membership function for fuzzy quantization. The preset evaluation level can be a predefined set of teaching effectiveness evaluation levels (such as excellent, good, qualified, unqualified), which can be used as the fuzzy set domain of the membership function output. The membership degree can be the degree to which the target evaluation parameter value belongs to a certain preset evaluation level, taking values ​​in the range [0, 1]. It can be used to convert precise parameter values ​​into fuzzy evaluation information, supporting subsequent reinforcement learning processing.

[0035] For each experimental course, a membership function corresponding to the target evaluation parameter is constructed. A fuzzy function mapping to a preset evaluation level can be designed for each target evaluation parameter. For example, this operation can be achieved by fitting a Gaussian membership function based on historical data distribution; in an exemplary embodiment, it can also be achieved by domain experts defining the inflection point parameters of a triangular or trapezoidal membership function, thereby achieving the technical effect of fuzzy representation of evaluation indicators and compatibility with nonlinearity and uncertainty. The target evaluation parameter values ​​corresponding to the experimental course are collected, and the actual values ​​of the target evaluation parameters can be extracted from experimental process records, sensor data, or manual scoring. Furthermore, this operation can be achieved through IoT device log collection, video behavior analysis, or digitization of teacher scoring sheets, thereby achieving the technical effect of obtaining the raw input for fuzzy quantification. The membership degree of the target evaluation parameter value within the preset evaluation level is calculated based on the membership function and the target evaluation parameter value. The target evaluation parameter value can be substituted into the corresponding membership function to output its membership degree at each evaluation level. In a specific embodiment, this operation can be achieved through a lookup table method or a real-time function calculation module, thereby achieving the technical effect of converting precise values ​​to fuzzy sets.

[0036] Step S600: Based on the parameter weight values ​​and membership degrees, the teaching effectiveness of the experimental course is dynamically optimized and evaluated through a reinforcement learning agent to determine the final teaching effect of the experimental course.

[0037] The reinforcement learning agent can be an autonomous decision-making model built based on reinforcement learning algorithms. It can dynamically adjust its evaluation strategy based on environmental feedback, enabling dynamic optimization and evaluation of the teaching effectiveness of experimental courses, and supporting continuous iteration and adaptive adjustment of the evaluation strategy. In one specific embodiment, the reinforcement learning agent can be one or more of the following: Q-learning-based agents, deep deterministic policy gradient (DDPG)-based agents, and proximal policy optimization (PPO)-based agents. The final teaching effect can be the evaluation result of the experimental course teaching effectiveness determined by the dynamic optimization of the reinforcement learning agent, which can be used as a basis for teaching improvement and decision support.

[0038] The teaching effectiveness of the experimental course is dynamically optimized and evaluated by a reinforcement learning agent based on parameter weights and membership degrees. The parameter weights and membership degrees can be combined as state inputs, and the reinforcement learning agent outputs an optimized evaluation strategy and calculates the final effect. In one exemplary embodiment, this operation can be achieved by the agent using a weighted membership vector as the state and adjusting the evaluation rules through a policy gradient method. Furthermore, it can also be achieved by the agent maintaining a value network, predicting the long-term teaching improvement benefits of different evaluation combinations, and selecting the optimal evaluation path, thereby achieving adaptive optimization and closed-loop learning in the evaluation process. To determine the final teaching effectiveness of the experimental course, the evaluation results output by the reinforcement learning agent can be integrated to generate an interpretable teaching effectiveness level or score. For example, this operation can be achieved by a post-processing module mapping the policy output to standard level labels or continuous scores, thereby achieving the technical effect of outputting a final evaluation conclusion that can be used for teaching feedback and strategy adjustment.

[0039] Taking the evaluation of experimental teaching in chemical engineering majors in universities as an example, the reinforcement learning-based experimental teaching effectiveness evaluation method in this embodiment can be as follows: In a chemical engineering experimental center of a university, the system identifies 20 experimental courses and extracts the laboratory ventilation level (teaching environment parameter), the distribution of students' prerequisite course grades (student characteristic parameter), and the reaction risk level (experimental characteristic parameter) of each course as influencing factors; the courses are divided into "high-risk synthesis experiment group", "basic physical property determination group" and "process simulation experiment group" through clustering; for the "high-risk synthesis experiment group", the system assigns a higher weight to "safe operation standardization" based on the analytic hierarchy process, and after preliminary evaluation, selects "emergency response speed", "protective equipment utilization rate" and "byproduct control accuracy" as target evaluation parameters; a triangular membership function is constructed for each course to map the actual observed values ​​to the "excellent / good / qualified / unqualified" level; the reinforcement learning agent receives the weighted membership vector as the state, obtains reward signals based on the historical teaching improvement effect (such as the decrease in accident rate and the number of student innovation proposals), dynamically adjusts the evaluation strategy, and finally outputs a more accurate teaching effectiveness evaluation result that reflects safety and innovation capabilities.

[0040] This embodiment provides a reinforcement learning-based method for evaluating the effectiveness of experimental teaching. It identifies several experimental courses within the teaching system and obtains influencing factors to form course groups. A structured evaluation framework is constructed by assigning weights to evaluation parameters using the analytic hierarchy process (AHP). Target evaluation parameters are selected based on group evaluation results to focus on key dimensions. Membership functions are constructed to fuzzy-quantify the target evaluation parameter values ​​to accommodate uncertainty. A reinforcement learning agent dynamically optimizes the evaluation using parameter weights and membership degrees as state inputs to output the final teaching effect. This method achieves a complete closed loop from context awareness, parameter selection, fuzzy quantification to intelligent optimization, ensuring that the evaluation results are both group-specific and individual-adaptive, while supporting the continuous evolution of teaching strategies. Thus, without relying on fixed rules, it improves the accuracy, fairness, scientific rigor, and practicality of experimental teaching effectiveness evaluation.

[0041] In one embodiment, several experimental courses are classified according to an impact factor to obtain multiple course groups consisting of at least one experimental course, specifically including: Based on the impact factor, determine the impact factor values ​​corresponding to several experimental courses.

[0042] The impact factor value can be a quantitative numerical representation of teaching environment parameters, student characteristic parameters, and experimental characteristic parameters in a specific experimental course. It can be used as a numerical basis for course classification and supports group division based on difference thresholds. In this embodiment, the impact factor value can be a numerical feature obtained by extracting raw information from the teaching system, student records, or experimental metadata and then standardizing or normalizing it. For example, the impact factor value can include, but is not limited to, one or more of the following: equipment configuration score, student prior knowledge level index, and experimental hazard level code value.

[0043] From a number of experimental courses, select designated experimental courses whose differences between the values ​​of each influencing factor are within a preset range to form a course group, thereby classifying the experimental courses.

[0044] The preset range can be an upper limit for the allowable difference in impact factor values ​​used to determine whether two or more experimental courses belong to the same course group. It can be used to control the balance between homogeneity within the course group and external discriminability, achieving adaptive clustering granularity. In an exemplary embodiment, the preset range can be dynamically set during the reinforcement learning training phase, using a larger value (relaxed threshold) initially and a smaller value (strict threshold) during the convergence period. Furthermore, the preset range can include, but is not limited to, Euclidean distance thresholds, Manhattan distance thresholds, and cosine similarity tolerance intervals. The designated experimental courses can be a group of experimental courses determined during the screening process to meet the requirement that the difference in impact factor values ​​is within the preset range, and can be used as the basic unit constituting a single course group.

[0045] Selecting specific experimental courses from a set of experimental courses whose differences between the corresponding impact factor values ​​are within a preset range can be achieved by calculating the differences between any two courses across each impact factor dimension. If all differences do not exceed the current preset range, they are grouped into the same group. Furthermore, this operation can be implemented by performing difference checks on all course combinations using a pairwise comparison method and merging courses that meet the conditions, or by using sliding windows or density clustering to include all courses within the neighborhood of a given course that meet the difference constraints. This allows for automatic course grouping based on numerical similarity, ensuring consistency of teaching context within each group. Classifying several experimental courses can be achieved by repeatedly executing the selection process until all experimental courses are assigned to at least one course group, thus completing the grouping organization of the entire course set.

[0046] The preset range is dynamically adjusted according to the reinforcement learning training stage. A more lenient threshold is used in the initial training stage, and a more stringent threshold is used in the convergence stage.

[0047] The reinforcement learning training phase can refer to different stages in the learning process of a reinforcement learning agent from an initial policy to a stable policy. It can be used as a trigger condition for dynamically adjusting a preset range, guiding the course group modeling from exploration to refinement. For example, the reinforcement learning training phase may include, but is not limited to, the initial training phase, the mid-term optimization phase, and the convergence phase.

[0048] The initial training phase can be the early learning stage where the reinforcement learning agent's policy is not yet stable and the reward signal fluctuates significantly. A relaxed threshold can be used in this stage to expand the coverage of course groups and enhance the model's ability to explore diverse teaching scenarios. The convergence phase can be the later learning stage where the reinforcement learning agent's policy tends to stabilize and the fluctuation of evaluation results significantly decreases. A strict threshold can be used in this stage to improve the consistency within course groups and enhance the sensitivity of evaluation to subtle differences. A relaxed threshold can be a relatively large preset range set in the initial training phase, allowing courses with large differences in influence factor values ​​to be grouped together. This can be used to promote the initial aggregation of cross-scenario courses and avoid the model getting stuck in local optima due to premature fine-grained grouping. A strict threshold can be a relatively small preset range set in the convergence phase, allowing only courses with highly similar influence factor values ​​to form groups. This can be used to improve the similarity of teaching scenarios within courses within a group, providing a high-fidelity modeling foundation for subsequent accurate evaluation.

[0049] Dynamically adjusting the preset range during the reinforcement learning training phase can monitor the training status of the reinforcement learning agent, such as the rate of change of the loss function and the variance of the evaluation results, and switch the preset range value strategy accordingly. Furthermore, this operation can be achieved by setting a fixed number of rounds for switching (using a loose threshold for the first N rounds, then switching to a strict threshold) or by adaptively adjusting based on performance metrics (automatically tightening the preset range when the evaluation stability exceeds the threshold for M consecutive rounds). This allows the course group construction process to evolve in tandem with the reinforcement learning optimization goal, balancing exploration and refinement. Using a looser threshold in the initial training phase means setting the preset range to a larger value in the first few rounds of reinforcement learning, allowing courses with greater differences to be grouped together, thereby expanding the diversity of group coverage and improving the model's initial generalization ability and policy exploration space. Using a stricter threshold in the convergence phase means narrowing the preset range after the reinforcement learning policy tends to stabilize, retaining only highly similar courses to form groups, thereby enhancing the homogeneity within groups and improving the accuracy of subsequent evaluation parameter selection and membership calculation.

[0050] For example, in the scenario of evaluating mechatronics training courses in vocational colleges, the reinforcement learning-based experimental teaching effectiveness evaluation method of this embodiment can be as follows: The system identifies 30 training courses, extracts the equipment age (teaching environment parameter), student programming foundation distribution (student characteristic parameter), and project integration complexity (experimental characteristic parameter) for each course, and quantifies them into three-dimensional influence factor values; In the early stage of reinforcement learning (the first 20 rounds), the preset range is set to 0.5 (relaxed threshold), and courses with newer equipment but weaker student foundation and courses with older equipment but stronger student foundation are grouped into the same "hybrid ability training group" to explore the teaching effect pattern under different resource configurations; As the training enters the convergence stage (after the 50th round), the preset range is tightened to 0.15 (strict threshold), at which point only courses with newer equipment, stronger student foundation, and more complex projects are grouped into the "advanced comprehensive training group," while basic operation courses are grouped separately; This dynamic grouping mechanism enables the reinforcement learning agent to widely sample different teaching modes in the early stage and focus on refined evaluation in the later stage, ultimately outputting evaluation results that are more in line with the actual teaching situation.

[0051] This embodiment achieves the technical effect of introducing a dynamic threshold mechanism driven by the reinforcement learning training phase on the basis of the original course group modeling: First, the influence factors are converted into calculable influence factor values, and then specified experimental courses are selected to form groups based on whether the difference is within a preset range. The key is that the preset range is adaptively adjusted with the reinforcement learning process. In the early stage, a loose threshold is used to expand the diversity of groups and enhance the model's exploration ability; in the convergence phase, a strict threshold is switched to improve the consistency within the groups and ensure the evaluation accuracy. This mechanism enables the course classification process and the reinforcement learning optimization goal to form a synergistic closed loop, which not only avoids the neglect of subtle differences in teaching contexts by traditional static clustering, but also overcomes the incompatibility of fixed thresholds at different stages of model training. Thus, it provides a more representative and stable group basis for weight allocation, fuzzy quantization and dynamic evaluation, and ultimately improves the accuracy, fairness, generalization ability and continuous optimization ability of the entire evaluation system.

[0052] In one embodiment, before selecting a specific experimental course from a plurality of experimental courses, the method further includes: Based on historical performance data from several experimental courses, corresponding learning behavior change curves are generated. Based on these learning behavior change curves, the current teaching stage of the experimental course is predicted using the reinforcement learning state recognition module. The range of differences between influencing factor values ​​is dynamically adjusted according to the teaching stage; the range of differences is positively correlated with the students' adaptability at the teaching stage: the lower the adaptability, the wider the range of differences.

[0053] Historical performance data can be a collection of quantitative results of student performance recorded in previous implementations of several experimental courses, which can be used as the original input for generating learning behavior change curves. In this embodiment, historical performance data is fitted with time series data such as scores, operation completion rate, and error rate of previous experiments in the same or similar courses, providing a foundation for constructing a continuous representation reflecting learning trends.

[0054] The learning behavior change curve can be a time-series function constructed based on historical performance data of several experimental courses, reflecting the evolution trend of students' abilities over time. It can be used to reveal the stage-specific characteristics of students' adaptation and ability development during the teaching process, providing data support for identifying teaching stages. Furthermore, the learning behavior change curve can be fitted with a continuous curve reflecting learning trends by performing time alignment and smoothing on the historical performance sequences of each experimental course. In an exemplary embodiment, the learning behavior change curve can use moving average or locally weighted regression (LOESS) to smooth discrete performance points, or it can use Gaussian process regression to model the uncertainty of performance changes with experimental rounds. Exemplarily, the learning behavior change curve can include, but is not limited to, one or more of the following: performance improvement curve, plateau stability curve, and fluctuating adaptation curve. The current experimental course can be an experimental teaching unit that is being evaluated or will soon be included in the course group, and can be used as an object for teaching stage prediction and difference range adjustment.

[0055] The reinforcement learning state recognition module can be a sub-model embedded in the reinforcement learning framework. It is used to determine the current stage of teaching based on input features, enabling intelligent perception of the current experimental course's teaching progress and driving dynamic adjustments to the difference range. In one specific embodiment, the reinforcement learning state recognition module can use features such as the slope, curvature, and volatility of the learning behavior change curve as input, and output teaching stage labels through a classification network or rule engine. For example, the reinforcement learning state recognition module can include, but is not limited to, one or more of the following: an LSTM-based state recognizer, a decision tree-based teaching stage classifier, and a clustering-based unsupervised state detector.

[0056] The teaching phase can be a time interval with specific learning objectives and student performance characteristics during the implementation of an experimental course, and can be used as a contextual basis for adjusting the granularity of course group division. For example, the teaching phase can include, but is not limited to, one or more of the following: initial adaptation period, skills consolidation period, and higher-order innovation period.

[0057] Generating corresponding learning behavior change curves based on historical performance data from several experimental courses can be achieved by time-aligning and smoothing the historical performance sequences of each experimental course to fit a continuous curve reflecting learning trends. Furthermore, this operation can be implemented by smoothing discrete performance points using moving averages or locally weighted regression (LOESS), or by using Gaussian process regression to model the uncertainty of performance changes with experimental rounds. This transforms discrete performance into an analyzable representation of the teaching process, supporting stage identification.

[0058] Based on the learning behavior change curve, the reinforcement learning state recognition module can predict the current teaching stage of the experimental course. This can be achieved by extracting key features of the learning behavior change curve (such as the first derivative, inflection point position, and variance) and inputting them into the state recognition module to output the teaching stage category. Furthermore, this operation can be implemented by training a lightweight neural network that takes curve segments as input and outputs probability distributions for three teaching stages, or by defining expert rules (if the curve slope is greater than a threshold and the variance is small, it is determined to be a skill consolidation period). This enables automatic perception of the teaching progress and provides contextual basis for dynamic grouping.

[0059] Student adaptability can refer to the level of familiarity and mastery of the current experimental course content, equipment, or procedures among the student group. It can be used as a moderating variable for setting the difference range and is positively correlated with it. The difference range can be the maximum allowable difference threshold for the influence factor values ​​used to determine whether two experimental courses can be grouped into the same group. It can be used to introduce a sense of teaching progress in course classification, enabling group modeling to consider both static attributes and dynamic development stages. In a specific embodiment, the difference range can be determined by the student adaptability corresponding to the teaching stage; the lower the adaptability, the wider the range. For example, the difference range can include, but is not limited to, one or more of the following: multidimensional Euclidean distance tolerance, weighted Manhattan distance upper limit, and multidimensional independent difference boundary.

[0060] The range of differences between influencing factor values ​​can be dynamically adjusted based on the teaching stage. This can be achieved by querying a mapping table between the teaching stage and the range of differences, or by calculating the appropriate range of differences using a function. This allows the granularity of course group division to adapt to the actual pace of teaching. Establishing a positive correlation between the range of differences and student adaptability at each teaching stage can be achieved by setting the range of differences as a monotonically increasing function of student adaptability, where adaptability is implicitly represented by the teaching stage. Furthermore, this operation can be implemented by pre-setting a fixed range of differences for each teaching stage (initial adaptation period → 0.6, skill consolidation period → 0.3, advanced innovation period → 0.15), or by constructing a continuous mapping function (e.g., the range of differences equals the proportional coefficient multiplied by one minus the normalized adaptability score). This ensures that group inclusivity is expanded when students' adaptability is weak, and the granularity of assessment is improved when their adaptability is strong.

[0061] For example, in the scenario of evaluating experimental courses in biomedical engineering majors in universities, the reinforcement learning-based experimental teaching effectiveness evaluation method in this embodiment can be as follows: The system collects the performance data of a certain biosignal acquisition experiment over the past 5 semesters and finds that students' performance increases slowly in the first 1-2 weeks (low slope, large fluctuations), rises rapidly in the 3-4 weeks, and tends to stabilize in the 5th week; the reinforcement learning state recognition module determines that the current period is the initial adaptation period, corresponding to a low degree of student adaptation; the system automatically sets the difference range between influencing factor values ​​(such as equipment complexity, prerequisite knowledge requirements, number of operation steps) to 0.55 (relatively wide), and groups this course with physiological signal processing experiments with slightly different equipment but similar teaching objectives into the same group to share the evaluation strategy; as the semester progresses, the learning curve enters the "skill consolidation period", and the difference range automatically tightens to 0.25. At this time, only highly similar courses are retained to form groups, so that the membership function and reinforcement learning evaluation focus more on subtle operational differences, thereby dynamically matching the teaching pace.

[0062] This embodiment transforms discrete scores into analyzable representations of the teaching process to support stage identification, enables automatic perception of the teaching process to provide contextual basis, allows the granularity of course group division to adapt to the actual teaching rhythm, and ensures that the group inclusion is expanded when students' adaptability is weak and the assessment precision is improved when their adaptability is strong. This achieves the goal of making course classification not only rely on static attribute similarity but also integrate the temporal evolution characteristics of the teaching process, solving the problem that traditional clustering methods ignore the learning development stages. The resulting course groups are more in line with the real teaching rhythm, providing a more timely and educationally reasonable modeling foundation for assessment parameter selection, membership quantification, and reinforcement learning optimization. Ultimately, it realizes the adaptive response of the assessment system to teaching dynamics, improving the fairness, guidance value, and robustness of the assessment.

[0063] In one embodiment, based on the group evaluation results, target evaluation parameters for evaluating the teaching effectiveness corresponding to the course group are selected from the evaluation parameters, specifically including: The fluctuation value between individual evaluation results corresponding to any two experimental courses in the group evaluation results is determined by the reinforcement learning policy network, and the largest fluctuation value is selected from the fluctuation values. If the maximum fluctuation value is within the preset range, the evaluation parameter will be used as the target evaluation parameter; Otherwise, the reinforcement learning action selection module is invoked to combine the evaluation parameters, generate an evaluation parameter set, and optimize the evaluation strategy through multiple rounds of iterative training based on the evaluation parameter set until the fluctuation values ​​corresponding to the parameter set all meet the preset range, at which point the evaluation parameter set is used as the target evaluation parameter.

[0064] The reinforcement learning policy network can be a neural network model that outputs the fluctuation values ​​of evaluation parameters across different experimental courses. It is a component of the reinforcement learning agent and can be used to quantify the consistency of evaluation parameters within a course group, providing a basis for selecting stable and effective target evaluation parameters. In this embodiment, the reinforcement learning policy network can receive each individual evaluation result from the group evaluation results as input and calculate the difference (i.e., fluctuation value) between any two course evaluation results through forward propagation. Furthermore, the reinforcement learning policy network can be one or more of the following: a fully connected policy network, a graph neural network, or an attention mechanism policy network.

[0065] The volatility value can be a measure of the difference between the single assessment results obtained when any two experimental courses in the same course group are initially assessed using the same assessment parameters. It can be used to reflect the stability and fairness of the assessment parameters within the group. For example, the volatility value can be one or more of the following, including but not limited to absolute difference volatility, relative percentage volatility, and standardized Euclidean distance volatility.

[0066] The maximum fluctuation value can be the maximum of the fluctuation values ​​of all pairs of courses in the group assessment results, and can be used as a key criterion to determine whether the current assessment parameter is suitable for the course group. The preset range can be a pre-set acceptable threshold range for fluctuation values, and can be used to determine whether the stability of the assessment parameter within the course group meets the standard.

[0067] Determining the fluctuation value between individual assessment results of any two experimental courses within a group assessment result using a reinforcement learning policy network can be achieved by inputting all individual assessment results within the group into the policy network, calculating the difference in assessment scores between each pair of courses, and outputting a fluctuation value matrix. Furthermore, this operation can be implemented through the forward propagation mechanism of the policy network, thereby enabling the automatic quantification of the consistency of assessment parameters across courses.

[0068] Filtering out the maximum fluctuation value from the fluctuation values ​​can be done by traversing the fluctuation value matrix and extracting the element with the largest value. Furthermore, this operation can be implemented using a numerical comparison algorithm to identify the course pair with the most unstable evaluation parameters in the group. Determining whether the maximum fluctuation value is within a preset range can be done by comparing the maximum fluctuation value with the upper and lower bounds of the preset range. Furthermore, this operation can be implemented using conditional logic to determine whether a parameter reorganization mechanism needs to be activated.

[0069] If the maximum fluctuation value is within a preset range, the evaluation parameter can be used as the target evaluation parameter. Alternatively, the current evaluation parameter can be retained, and the process can directly proceed to the subsequent membership calculation stage. Furthermore, this operation can be achieved through a parameter locking mechanism, thereby ensuring that only parameters that meet the stability criteria are used for the final evaluation.

[0070] The reinforcement learning action selection module can be a submodule within a reinforcement learning agent responsible for generating new combinations of evaluation parameters. It can be used to automatically construct more group-fit ​​composite evaluation parameters when the stability of the original evaluation parameters is insufficient. In an exemplary embodiment, the reinforcement learning action selection module can generate a new set of evaluation parameters by selecting parameter combination strategies (such as weighted fusion, logical AND / OR, nonlinear transformation, etc.) from the action space based on the performance feedback of the current evaluation parameters (e.g., fluctuation values ​​exceeding limits). Exemplarily, the reinforcement learning action selection module can include, but is not limited to, one or more of discrete action selectors, continuous parameter synthesizers, and hybrid policy action modules.

[0071] When the maximum fluctuation value exceeds the preset range, the reinforcement learning action selection module is invoked to combine the evaluation parameters and generate an evaluation parameter set. This can be achieved by triggering the action selection module to select parameter combinations from a predefined action space to generate a new evaluation parameter set. Furthermore, this operation can generate new parameters by linearly weighting the original evaluation parameters, or by constructing composite indicators through logical operations (such as "operational standardization and teamwork"), thereby constructing a new evaluation dimension that better reflects the commonalities of the course group.

[0072] The evaluation parameter set can be a novel set of multidimensional evaluation indicators generated by combining the original evaluation parameters. It can be used to replace unstable single evaluation parameters, improving the consistency and discriminative power of evaluations within a group. In one specific embodiment, the evaluation parameter set can be formed by performing algebraic or logical operations on the original evaluation parameters through a reinforcement learning action selection module. For example, the evaluation parameter set can include, but is not limited to, one or more of the following: a weighted fusion parameter set, a Boolean logic combination parameter set, and a nonlinear mapping parameter set.

[0073] Optimizing the evaluation strategy through multiple rounds of iterative training based on the evaluation parameter set can be achieved by re-executing the initial evaluation with a new parameter set, calculating new fluctuation values, and using the degree of reduction in fluctuation values ​​as a reward signal to update the policy network and action selection module. Furthermore, this operation can be implemented by using policy gradient methods to optimize the parameter combination strategy end-to-end, or by using Q-learning to update the action value function to select a better combination, thereby enabling the self-correction and continuous evolution of the evaluation system.

[0074] To determine whether the fluctuation values ​​corresponding to the parameter set all meet the preset range, one can recalculate the fluctuation values ​​of all pairs of courses after using the new parameter set and verify whether their maximum values ​​fall within the preset range. Furthermore, this operation can be achieved by recalculating the fluctuation value matrix and comparing it with the threshold, thereby confirming that the parameter set has met the requirements for consistency in group evaluation.

[0075] When all fluctuation values ​​meet a preset range, the evaluation parameter set is used as the target evaluation parameter. This can be achieved by formally designating the converged evaluation parameter set as the target evaluation parameter for subsequent membership function construction. Furthermore, this operation can be implemented through a parameter set solidification mechanism, enabling dynamic transformation from original parameters to group-fitting parameters. The evaluation strategy can be a set of decision rules adopted by the reinforcement learning agent during the selection and optimization of target evaluation parameters. It can guide the direction of parameter combination and iterative training to minimize evaluation fluctuations within the group.

[0076] For example, in the scenario of evaluating and optimizing mechatronics training courses in vocational colleges, the reinforcement learning-based experimental teaching effectiveness evaluation method in this embodiment can be as follows: 12 mechatronics training courses in a vocational college are clustered into a "high equipment dependence group". After initial weighted evaluation using evaluation parameters including "wiring accuracy", "troubleshooting time", and "safety procedure compliance rate", the reinforcement learning strategy network calculation found that the "troubleshooting time" fluctuates greatly between CNC maintenance and PLC debugging courses (up to 0.45), exceeding the preset range [0, 0.2]. The system then calls the reinforcement learning action selection module to attempt to nonlinearly normalize and combine "troubleshooting time" and "equipment complexity coefficient" to generate a new parameter "standardized troubleshooting efficiency". After three rounds of iterative training, the maximum fluctuation value of the new parameter set is reduced to 0.18, meeting the requirements, and is therefore used as the target evaluation parameter. Subsequent membership functions are constructed based on this new parameter, making the evaluation results more fairly reflect the students' true abilities under different equipment conditions.

[0077] This embodiment provides a reinforcement learning-based method for evaluating the effectiveness of experimental teaching. By utilizing a reinforcement learning strategy network to quantify the volatility of evaluation parameters within a course group, if the maximum volatility value exceeds a preset range, the reinforcement learning action selection module is activated to intelligently combine the original evaluation parameters to generate an evaluation parameter set. Through multiple rounds of iterative training, the evaluation strategy is optimized with the goal of converging the volatility value until the new parameter set exhibits sufficient stability within the group. This transforms the selection of target evaluation parameters from static preset to dynamic generation, solving the cross-course unfairness problem caused by fixed parameters in traditional methods. It endows the evaluation system with self-diagnosis and self-repair capabilities, ensuring that the final selected target evaluation parameters are highly consistent with the teaching characteristics of the course group. This significantly improves the internal consistency, cross-course comparability, and long-term adaptability of the evaluation results, further enhancing the scientific nature and intelligent evolution capabilities of the entire evaluation method.

[0078] In one embodiment, after iterative evaluation via reinforcement learning, the method further includes: When multiple sets of evaluation parameters have fluctuation values ​​that meet the preset range, calculate the cumulative reward value for each set of parameters. Based on the cumulative reward value, the set of evaluation parameters with the highest reward value is selected as the target evaluation parameter; the cumulative reward value is calculated by the reinforcement learning agent based on the improvement of teaching effect and evaluation stability.

[0079] The cumulative reward value can be a comprehensive feedback signal accumulated by the reinforcement learning agent for a certain set of evaluation parameters in multiple rounds of iterative evaluation. It can be used as the basis for selecting the best among multiple qualified evaluation parameter sets, achieving a decision that balances the guiding force of teaching improvement and the robustness of evaluation. In this embodiment, the cumulative reward value can be calculated by the reinforcement learning agent based on two dimensions: the magnitude of improvement in teaching effectiveness and the stability of evaluation. Furthermore, the cumulative reward value can include, but is not limited to, one or more of the following: cumulative reward primarily driven by teaching improvement, cumulative reward primarily driven by stability, and balanced cumulative reward.

[0080] The improvement in teaching effectiveness can be defined as the observable improvement in teaching quality during subsequent teaching cycles after adopting a certain set of evaluation parameters. It can reflect the actual effectiveness of the evaluation parameter set in positively guiding teaching. In an exemplary embodiment, the improvement in teaching effectiveness can be quantified by comparing changes in student ability indicators (such as operational proficiency, number of innovative proposals, and reduction in error rate) before and after using the parameter set. For example, the improvement in teaching effectiveness may include improvements in the depth of knowledge acquisition, the efficiency of practical operations, and the quality of teamwork.

[0081] Evaluation stability refers to the consistency of evaluation results output by a set of evaluation parameters across different time points or different subgroups of courses. It can be used to measure the robustness and generalization reliability of the evaluation parameter set. In one specific embodiment, evaluation stability can be measured by calculating the standard deviation or coefficient of variation of the fluctuation values ​​of the parameter set across multiple batches of evaluations. Furthermore, evaluation stability can include cross-time period stability, cross-subgroup stability, and noise interference resistance stability.

[0082] Weighted calculation can be a mathematical operation that linearly or nonlinearly fuses the improvement in teaching effectiveness with the stability of evaluation according to preset or learnable weights. It can be used to achieve preference balance in multi-objective optimization and support value-oriented parameter set selection. In this embodiment, weighted calculation can be implemented by having the weights set by experts or by adaptive adjustment by the reinforcement learning agent. The fusion methods include weighted summation, product normalization, etc.

[0083] When multiple evaluation parameter sets have fluctuation values ​​that meet a preset range, the cumulative reward value for each parameter set is calculated. This can be achieved by, after multiple evaluation parameter sets with all fluctuation values ​​meeting the criteria are generated during the screening phase, calling the reinforcement learning agent to backtrack its historical interaction records and summarizing the long-term reward signal corresponding to each set of parameters. Furthermore, this operation can improve the accuracy of parameter selection by identifying the candidate with the best overall performance from multiple candidate solutions with acceptable stability.

[0084] The reinforcement learning agent calculates the cumulative reward value based on the improvement in teaching effectiveness and the stability of evaluation. This can be achieved by extracting the changes in teaching improvement indicators and the consistency indicators of evaluation output triggered by each set of evaluation parameters during training, and then fusing them according to their weights to generate a single cumulative reward value. Furthermore, this operation can be achieved by using a fixed weight (e.g., 0.6:0.4) for linear weighted summation, or by introducing a gating mechanism to dynamically adjust the contribution ratio of the two, thereby enabling a joint evaluation of the "usefulness" and "reliability" of the evaluation parameter set.

[0085] The evaluation parameter set with the highest cumulative reward value is selected as the target evaluation parameter. This can be achieved by comparing the cumulative reward values ​​of all qualified evaluation parameter sets and selecting the parameter set with the highest value to proceed to the subsequent membership function construction process. Furthermore, this operation ensures that the final selected target evaluation parameter is not only stable but also has the greatest promoting effect on improving teaching quality.

[0086] For example, in the scenario of optimizing and evaluating a group of biomedical engineering experimental courses in universities, the reinforcement learning-based experimental teaching effectiveness evaluation method in this embodiment can be as follows: In the "medical instrument operation experimental group," after parameter recombination and iterative training, the system generates three sets of evaluation parameters, all with fluctuation values ​​less than 0.18 (preset upper limit 0.2): Set A emphasizes "equipment calibration accuracy," Set B integrates "emergency response + data recording integrity," and Set C adopts the product form of "operational fluency × safety compliance rate." The reinforcement learning agent reviews data from the past two semesters and finds that classes using Set B have an average score increase of 12% in subsequent clinical simulation assessments, with a standard deviation of only 0.05 between the three assessments; while Set C, although showing an increase of 15%, has poor stability (standard deviation 0.11). The system calculates the cumulative reward with weights of 0.7 (increase) and 0.3 (stability). Set B has the highest score and is therefore selected as the target evaluation parameter. Subsequent teaching is then based on this to strengthen emergency response and recording training, forming a virtuous cycle.

[0087] This embodiment provides a reinforcement learning-based experimental teaching effectiveness evaluation method. It calculates the cumulative reward value of each parameter set when the fluctuation values ​​of multiple evaluation parameter sets meet a preset range, and selects the evaluation parameter set with the highest cumulative reward value as the target evaluation parameter. The cumulative reward value is calculated by the reinforcement learning agent based on the improvement in teaching effectiveness and the stability of the evaluation. By introducing a joint optimization mechanism based on the effectiveness of teaching improvement and the consistency of evaluation from multiple candidate parameter sets that meet stability standards, the method upgrades the selection of evaluation parameters from a single stability standard to multi-objective value optimization. This ensures that the evaluation system not only has internal consistency but also maximizes its contribution to improving teaching quality. It achieves a functional leap from describing the status quo to driving progress in the evaluation model, effectively overcoming the shortcomings of static parameter fixation and the lack of a closed-loop teaching feedback mechanism in traditional methods. This further enhances the scientific rigor, fairness, and practicality of the evaluation results and provides a highly reliable basis for the precise adjustment of teaching strategies.

[0088] In one embodiment, constructing the membership function corresponding to the target evaluation parameters specifically includes: The evaluation levels corresponding to the target evaluation parameters are divided; wherein each evaluation level is mapped to a state space interval in reinforcement learning. This operation can be achieved by semantically dividing the range of values ​​for the target evaluation parameters into several mutually exclusive and fully covered evaluation level intervals. Furthermore, this operation can be implemented by setting semantic levels such as "excellent," "good," "qualified," and "unqualified" as continuous or discrete numerical sub-intervals, thus providing a semantic basis for subsequent mapping to the state space and membership optimization.

[0089] In reinforcement learning, the state space interval can be a set of mutually exclusive intervals that divide continuous or discrete states, with each interval corresponding to a semantically defined evaluation level. In this embodiment, the state space interval can map a preset evaluation level to a numerical interval in the reinforcement learning state space, serving as the state input unit for Q-learning. For example, the state space interval can be one or more of high-performance, medium-performance, and low-performance state intervals. Further, mapping each evaluation level to a state space interval in reinforcement learning can be achieved by assigning a corresponding numerical interval identifier in the reinforcement learning state space to each evaluation level. In an exemplary embodiment, this mapping can be implemented by mapping "excellent" to a continuous interval such as [0.8, 1.0] and "good" to [0.6, 0.8), or by assigning a discrete integer state ID (such as 1, 2, 3…) to each level, thereby establishing a formal correspondence between the fuzzy evaluation and the reinforcement learning state representation.

[0090] For each target evaluation parameter, the Q-learning algorithm is used to optimize the level membership degree of the parameter in each level interval.

[0091] Q-learning can be a model-free reinforcement learning algorithm that learns an optimal policy by maintaining a state-action value function (Q-table) and iteratively updating it based on the Bellman equation. In a specific embodiment, Q-learning allows the agent to select actions based on the current state in a discrete state and action space, update its Q-value after receiving environmental rewards, and eventually converge to the optimal policy. Q-learning can be used to dynamically optimize the membership degree of the target evaluation parameters in each evaluation level interval, enabling the membership function to have adaptive capabilities based on teaching feedback. For example, Q-learning can employ one or more of the following: tabular Q-learning, deep Q-network (DQN), and dual Q-learning.

[0092] The membership degree can be defined as the degree to which a target evaluation parameter value belongs to a specific evaluation level, and in this scheme, it serves as an adjustable parameter for Q-learning optimization. In this embodiment, the membership degree can be initially set by experts or generated through data fitting, and then iteratively adjusted by Q-learning based on reward signals. The membership degree can be one or more of the following: excellent membership degree, good membership degree, and acceptable membership degree. Furthermore, optimizing the membership degree of the parameter in each level interval using the Q-learning algorithm can be achieved by treating the membership degree as a learnable parameter, using the state space interval as the state, and the membership degree adjustment action as the action, updating the Q value and optimizing the membership degree based on reward feedback. In an exemplary embodiment, this optimization can be achieved by modeling the membership degree adjustment as a fine-tuning action on the inflection point of the membership degree function, with Q-learning selecting the optimal adjustment direction, or by directly outputting each membership degree as the Q value, selecting the level corresponding to the maximum Q value as the current evaluation result through a strategy, and updating the membership degree in reverse, thereby enabling the membership degree function to dynamically evolve according to actual teaching feedback, thus breaking away from static prior settings.

[0093] The membership degree of the level is re-standardized based on reward feedback.

[0094] The reward feedback can be a scalar signal from the teaching assessment environment, used to indicate the degree of consistency between the assessment results and the actual teaching effectiveness under the current membership configuration. In this embodiment, the reward feedback can be derived from the teaching improvement effects (such as improved student performance, reduced experimental accidents, teacher satisfaction, etc.) and used as the reward input for Q-learning. The reward feedback can be one or more of the following: positive rewards for teaching effectiveness, penalty signals for assessment deviations, and cumulative rewards for long-term teaching improvement.

[0095] Re-standardization can be achieved by normalizing the membership degrees of each level after Q-learning optimization, ensuring that the sum of membership degrees in a fuzzy set is 1 or other constraints. In a specific embodiment, re-standardization can employ weighted normalization or soft maximization methods, forcibly satisfying the normative requirements of fuzzy mathematics while preserving the optimization direction. Furthermore, re-standardization of membership degrees based on reward feedback can be performed after Q-learning updates, normalizing the membership degrees of all levels to satisfy the constraints of fuzzy set theory. For example, this re-standardization can be achieved by using a softmax function to smoothly normalize the optimized membership degrees, or by scaling proportionally to make the sum of membership degrees equal to 1, while preserving relative size relationships. This ensures the mathematical validity and interpretability of the membership function while maintaining the optimization results of reinforcement learning.

[0096] Obtain the membership function corresponding to the target evaluation parameters.

[0097] This operation can integrate the re-standardized membership degrees at each level to form a complete piecewise or continuous membership degree function, thereby outputting a dynamic membership degree function that combines data-driven adaptability with fuzzy logic normalization.

[0098] For example, in the scenario of evaluating mechatronics training courses in higher vocational colleges, the reinforcement learning-based experimental teaching effectiveness evaluation method of this embodiment can be as follows: In a certain training center, for the target evaluation parameter of "PLC control wiring standardization", the system divides its evaluation level into four levels: "excellent", "good", "qualified" and "unqualified", and maps them to four intervals in the reinforcement learning state space: [0.75, 1.0), [0.5, 0.75), [0.25, 0.5), and [0, 0.25). The initial membership degree is set by experts, and then the Q-learning agent evaluates the students' membership degree based on the results of each round of training. The system generates reward feedback based on indicators such as fault diagnosis efficiency, teacher scoring consistency, and equipment damage rate, dynamically adjusting the membership values ​​for each level. For example, when it is found that students in the original "good" range perform close to the "excellent" group in subsequent tasks, Q-learning increases the tendency of that range to shift towards "excellent," adjusting the membership value accordingly. After each update, the system performs softmax re-standardization on the membership values ​​of the four levels to ensure that their sum is 1. The resulting membership function can more accurately reflect the actual ability distribution of the school's students in the training project, supporting the reinforcement learning agent to make fairer evaluation decisions in the future.

[0099] This embodiment provides a reinforcement learning-based method for evaluating the effectiveness of experimental teaching. By constructing a membership function corresponding to the target evaluation parameters, dividing the evaluation levels of the target evaluation parameters into semantic intervals and mapping them to the state space intervals of reinforcement learning, and using the Q-learning algorithm to dynamically optimize the membership degrees corresponding to each evaluation level under the guidance of reward feedback, and then re-standardizing the optimized membership degrees based on reward feedback, the membership function can be transformed from a static prior setting into a dynamic component with the ability to adapt to the teaching environment. This significantly enhances the evaluation model's ability to characterize complex and variable experimental teaching scenarios and its generalization performance, thereby supporting the entire evaluation system to achieve high-precision, adaptive, and interpretable dynamic optimization of the evaluation effect.

[0100] In one embodiment, determining the final teaching effect corresponding to the experimental course specifically includes: An evaluation vector is generated based on the membership degree of the target evaluation parameter value within each evaluation level.

[0101] The evaluation vector can be a vector composed of the membership degrees of a single target evaluation parameter across all preset evaluation levels. This vector can be used to preserve the complete distribution information of the parameter in the fuzzy level space, serving as the basic building block of the state feature matrix. Generating the evaluation vector based on the membership degrees of the target evaluation parameter value within each evaluation level can be achieved by arranging the membership degrees of a single target evaluation parameter across all preset evaluation levels into a one-dimensional vector in a fixed order. Furthermore, this operation enables a structured expression of the fuzzy evaluation information of the parameter, preserving the relative relationships between levels.

[0102] The evaluation vectors of all target evaluation parameters are concatenated into a state feature matrix, which is then input into the reinforcement learning evaluation model.

[0103] The state feature matrix can be a structured input matrix formed by concatenating the evaluation vectors of all target evaluation parameters. As the state representation of the reinforcement learning evaluation model, it can be used to uniformly encode multi-dimensional fuzzy evaluation information, providing the policy network with a complete and structured input for understanding the teaching context. In an exemplary embodiment, the state feature matrix can be a row-based parameter concatenation matrix, a column-based rank-aligned matrix, or a graph-structured feature matrix, etc. Concatenating the evaluation vectors of all target evaluation parameters into the state feature matrix can be done by concatenating the evaluation vectors horizontally or vertically in a preset order of the target evaluation parameters to form a two-dimensional matrix. For example, this operation can be achieved by treating each evaluation vector as a row to form a parameter × rank matrix, or by flattening all membership degrees into long vectors and reshaping them into a square matrix to adapt to the convolutional network, thereby constructing a unified state input that the reinforcement learning model can process, integrating multi-dimensional fuzzy evidence. Inputting the state feature matrix into the reinforcement learning evaluation model can be done by sending the concatenated state feature matrix into a trained or currently being trained reinforcement learning evaluation model for forward inference, thereby initiating the intelligent evaluation process and triggering the decision calculation of the policy network.

[0104] The policy network of the reinforcement learning evaluation model is subjected to fuzzy transformation and reward update to output a confidence vector of teaching effectiveness.

[0105] The reinforcement learning evaluation model can be a reinforcement learning architecture specifically designed for evaluating the effectiveness of experimental teaching. It integrates state input, policy decision-making, and reward mechanisms, enabling intelligent fusion and dynamic decision-making of multi-dimensional fuzzy evaluation information, replacing traditional static weighted rules. In a specific embodiment, the reinforcement learning evaluation model can include, but is not limited to, one or more of the following: Actor-Critic-based evaluation models, DQN-based evaluation models, and PPO-based evaluation models. The policy network can be a neural network module in the reinforcement learning evaluation model responsible for mapping from states to actions or action probability distributions. It can be used to perform high-order fusion of fuzzy information and decision generation, supporting the adaptive evolution of evaluation policies. Furthermore, the policy network can be a deterministic policy network, a stochastic policy network, or an attention-enhanced policy network. The fuzzy transformation can be a nonlinear mapping operation performed by the policy network on the input state feature matrix. It is used to fuse multi-source fuzzy information and extract high-level semantic features, going beyond linear weighting to capture the interaction effects and nonlinear correlations between evaluation parameters, thereby improving evaluation discriminative power.

[0106] In an exemplary embodiment, fuzzy transformation can be achieved through neural network layers (such as fully connected layers or attention mechanisms) to perform weighted aggregation, feature crossing, or context-aware membership reorganization. Reward updates can be a process of adjusting policy network parameters based on the consistency feedback between teaching effectiveness evaluation results and actual teaching effectiveness. This can be used to drive the policy network to continuously optimize the evaluation strategy, achieving closed-loop learning and adaptive evolution of the evaluation model. Furthermore, reward updates can utilize subsequent teaching indicators (such as student ability improvement, experiment reproduction success rate, and teacher feedback) to construct reward signals, and update the policy network weights through backpropagation. The teaching effectiveness confidence vector can be a probability distribution vector output by the policy network, representing the confidence level of each possible teaching effectiveness level. This can be used to provide an interpretable probabilistic basis for the final teaching effectiveness judgment, supporting uncertainty quantification and risk-aware decision-making.

[0107] For example, the teaching effectiveness confidence vector can be a discrete-level confidence vector, a continuous effect score distribution, or a multi-granularity hybrid confidence representation. By performing fuzzy transformation and reward update through the policy network of the reinforcement learning evaluation model, the policy network can perform a nonlinear mapping on the state feature matrix to generate the confidence vector and update the network parameters based on the reward signal from the environment. Furthermore, this operation can be achieved by dynamically weighting the membership degrees of different evaluation parameters using an attention mechanism before fusion, or by using LSTM to model the evaluation vector sequence to capture the temporal dependencies between parameters (suitable for multi-stage experiments), thereby achieving intelligent abstraction and continuous optimization of the evaluation logic, surpassing the expressive power of manual rules. The output teaching effectiveness confidence vector can be a probability distribution vector output by the policy network, where each element corresponds to the confidence level of a teaching effectiveness level, thus providing an interpretable and quantifiable representation of the evaluation results and supporting uncertainty perception.

[0108] The final teaching effect of the experimental course is determined based on the optimal action selected by the reinforcement learning agent in the teaching effect confidence vector.

[0109] The optimal action can be the teaching effect judgment behavior chosen by the reinforcement learning agent in the current state, which is expected to maximize the long-term cumulative reward. It can be used as the decision basis for the final teaching effect, corresponding to the evaluation level with the highest confidence in the confidence vector or its weighted selection result. Determining the final teaching effect of the experimental course based on the optimal action chosen by the reinforcement learning agent in the teaching effect confidence vector can be achieved by selecting the highest confidence level based on the confidence vector, or by selecting actions based on probability sampling, as the final evaluation conclusion. For example, this operation can be achieved by selecting the level corresponding to the argmax confidence as the final effect, or by introducing a threshold mechanism to trigger manual review when the highest confidence level falls below the threshold, thereby completing the decision transformation from a fuzzy confidence distribution to a deterministic evaluation conclusion.

[0110] For example, in the scenario of clinical skills simulation experiment assessment in medical schools, the reinforcement learning-based experimental teaching effectiveness evaluation method of this embodiment can be as follows: In a certain medical school OSCE (Objective Structured Clinical Examination) system, for the target assessment parameter of "standardization of cardiopulmonary resuscitation operation", the system has calculated its membership degree in the four levels of "excellent", "good", "qualified" and "unqualified" as [0.3, 0.5, 0.2, 0.0], forming an assessment vector; similarly, "team communication efficiency" and "accuracy of emergency judgment" also generate their own assessment vectors; the three are concatenated into a 3×4 state feature. The feature matrix is ​​input into the reinforcement learning evaluation model trained based on the PPO algorithm. The policy network assigns higher weights to "accuracy of emergency judgment" through an attention mechanism, because it is strongly correlated with patient survival rate in historical data. After performing fuzzy transformation, it outputs a teaching effect confidence vector [0.1, 0.65, 0.2, 0.05]. The agent selects "good" as the optimal action. If the students in this group perform well in real emergency drills, the system will give positive rewards, prompting the policy network to be more inclined to map similar membership combinations to "good" or "excellent" in the future, thereby achieving continuous calibration of the evaluation strategy.

[0111] This embodiment achieves the following technical effects: it preserves the complete membership distribution of each target evaluation parameter in the grade space through structured preservation; it uniformly encodes multi-parameter fuzzy evidence; it performs fuzzy transformations such as nonlinear fusion and attention weighting, and updates rewards in conjunction with teaching practice feedback; it provides probabilistic and interpretable evaluation basis and completes the decision transformation from confidence distribution to deterministic conclusion; it avoids the subjective limitations of manually setting evaluation rules; it achieves continuous optimization of evaluation strategies through the exploration-exploitation mechanism of reinforcement learning; it enables the system to adapt to changes in the teaching environment, student group evolution, and experimental project iteration; it significantly improves the accuracy, robustness, and interpretability of evaluation; and it supports the intelligent closed-loop feedback of teaching management.

[0112] In one embodiment, the final teaching effect of the experimental course is determined based on the optimal action selected by the reinforcement learning agent in the teaching effect confidence vector, specifically including: The evaluation level with the highest probability value is selected from the confidence vector of teaching effectiveness by the reinforcement learning agent.

[0113] The selection probability value can be the maximum probability value corresponding to a certain evaluation level in the teaching effectiveness confidence vector, reflecting the agent's confidence in the judgment and serving as a decision-making basis for whether to adopt the current evaluation conclusion. In this embodiment, the selection probability value is obtained by extracting the maximum value from the probability distribution output by the reinforcement learning agent based on the current state feature matrix. Furthermore, this operation can be achieved by taking the evaluation level corresponding to the element with the largest value in the teaching effectiveness confidence vector, thereby achieving the technical effect of determining the current optimal evaluation candidate.

[0114] If the selected probability value is lower than the preset threshold, the exploration mechanism is triggered to regenerate the evaluation parameter set and return to the iterative evaluation step.

[0115] The preset threshold can be a pre-defined confidence threshold used to determine whether the current evaluation result is reliable enough to be used as a final conclusion. It can be used to control the balance between exploration and utilization, and prevent low-confidence decisions from becoming fixed. In this embodiment, the preset threshold is dynamically configured by the system based on historical evaluation stability and misjudgment tolerance, or set manually.

[0116] The exploration mechanism can be an adaptive strategy that proactively adjusts the evaluation parameter configuration and re-executes the evaluation process when the reinforcement learning agent lacks confidence in the current evaluation result. It can be used to avoid outputting unreliable conclusions when evidence is insufficient or the context is abnormal, thereby improving the system's robustness and introspection capabilities under uncertain conditions. In an exemplary embodiment, the exploration mechanism can be activated when the selection probability value is lower than a preset threshold. It generates a new set of evaluation parameters by perturbing, expanding, or replacing some of the target evaluation parameters and returns to the group evaluation or membership calculation stage. Exemplarily, the exploration mechanism can include, but is not limited to, one or more of the following: parameter perturbation exploration, new dimension exploration, and course group re-clustering exploration.

[0117] The evaluation parameter set can be a new set of target evaluation parameters regenerated after the exploration mechanism is triggered, used for a new round of evaluation. This can provide new evaluation perspectives or dimensions to address situations where the original parameter set cannot effectively distinguish teaching effectiveness. The iterative evaluation step can start by regenerating the evaluation parameter set and re-executing the complete process of membership calculation, state feature construction, and reinforcement learning evaluation. This can be used to achieve closed-loop retrying and optimization of the evaluation process. Furthermore, this operation can activate the exploration mechanism, generate a new evaluation parameter set, and jump to the membership function construction or group evaluation stage for re-execution, thereby proactively seeking better evaluation configurations in low-confidence scenarios and avoiding misjudgments. In a specific embodiment, this includes randomly replacing some target evaluation parameters while retaining high-contribution parameters, introducing new dimensions (such as student emotional state and device response latency) to expand the evaluation parameter set, and performing fine-grained re-clustering of the original course groups to generate a more suitable parameter subset.

[0118] Otherwise, the assessment level will be used as the final teaching outcome, and the assessment results will be stored in the experience replay pool to optimize subsequent strategies.

[0119] The evaluation results can be a complete evaluation record that includes the final teaching effect, confidence vector, whether exploration was triggered, original parameter set and reward signal. It can be used as a storage unit for the experience replay pool to support subsequent strategy optimization.

[0120] An experience replay pool can be a buffer structure used to store historical evaluation interaction data (including states, actions, rewards, and next states). This can break data correlations, improve sample utilization efficiency, and support stable and efficient updates of the policy network. In an exemplary embodiment, after each determination of the final teaching effect, the experience replay pool can store the complete evaluation trajectory (including the state feature matrix, selected actions, reward feedback, whether exploration was triggered, etc.) in the pool for offline batch training of the policy network. Exemplarily, the experience replay pool can include, but is not limited to, one or more of the following: a priority experience replay pool, a hierarchical experience replay pool, and a course group-specific replay pool.

[0121] Subsequent strategies can be the evaluation decision strategies updated by the reinforcement learning evaluation model after receiving new experiences, reflecting the improvement logic learned by the system from historical evaluations. Furthermore, this operation can be achieved by storing the complete trajectory of this evaluation (state, action, reward, whether to explore, etc.) in an experience replay pool for periodic sampling and training by the policy network, thereby achieving long-term accumulation of evaluation experience and continuous evolution of the strategy. In a specific embodiment, this includes employing priority experience replay, assigning higher sampling weights to high-error or high-reward samples, or storing samples in partitions by course groups to support group-specific policy fine-tuning.

[0122] Taking the evaluation of interdisciplinary innovation experimental projects in new engineering disciplines as an example, the reinforcement learning-based experimental teaching effect evaluation method in this embodiment can be as follows: A university is conducting an experiment on "Intelligent Agricultural Robot System Integration" for the first time. The initial evaluation parameter set includes "hardware debugging efficiency," "algorithm robustness," and "team collaboration fluency." The teaching effect confidence vector output by the reinforcement learning agent is [0.35, 0.4, 0.2, 0.05]. The maximum selection probability value of 0.4 is lower than the preset threshold of 0.6, and the system determines that the current evidence is insufficient for reliable evaluation. An exploration mechanism is then triggered: a new parameter, "interdisciplinary knowledge integration depth," is introduced, and the weights of the original parameters are adjusted to form a new evaluation parameter set. The system returns to the previous step to re-select the target parameters and construct the membership function. The second round of evaluation yields a confidence vector of [0.1, 0.2, 0.65, 0.05]. At this point, the maximum probability of 0.65 exceeds the threshold, and the final teaching effect is determined to be "qualified." The complete assessment trajectory (including two attempts, parameter changes, and teacher supplementary feedback as a reward) was stored in the experience playback pool; in subsequent similar interdisciplinary experimental assessments, the strategy network, having learned this pattern, can more quickly identify the need to introduce knowledge integration dimensions, significantly improving the initial accuracy of the new curriculum assessment.

[0123] This embodiment provides a method for evaluating the effectiveness of experimental teaching based on reinforcement learning. It selects the evaluation level with the highest probability of selection by the reinforcement learning agent from the teaching effectiveness confidence vector. It then determines whether this selection probability is below a preset threshold to decide whether to trigger an exploration mechanism. If triggered, the evaluation parameter set is regenerated and the iterative evaluation step is returned. Otherwise, the evaluation level is used as the final teaching effect, and the evaluation results are stored in an experience replay pool to optimize subsequent strategies. By dynamically switching between utilization and exploration modes based on confidence level, actively reconstructing evaluation parameters and restarting the evaluation process at low confidence levels, and solidifying conclusions and accumulating experience for strategy updates at high confidence levels, the evaluation system can achieve self-reflection and proactive optimization capabilities. This significantly improves robustness, adaptability, and long-term evolutionary ability when facing new courses, new student groups, or sudden changes in the teaching environment, ensuring that the evaluation is both accurate and has continuous evolutionary potential.

[0124] Furthermore, to achieve the above objectives, the present invention also provides an experimental teaching effectiveness evaluation device based on reinforcement learning. The device includes: a memory, a processor, and an experimental teaching effectiveness evaluation program based on reinforcement learning stored in the memory and executable on the processor. The experimental teaching effectiveness evaluation program based on reinforcement learning is configured to implement the steps of the experimental teaching effectiveness evaluation method based on reinforcement learning as described above.

[0125] In addition, to achieve the above objectives, the present invention also provides a medium storing a reinforcement learning-based experimental teaching effectiveness evaluation program, wherein when the reinforcement learning-based experimental teaching effectiveness evaluation program is executed by a processor, it implements the steps of the reinforcement learning-based experimental teaching effectiveness evaluation method as described above.

[0126] Other embodiments or specific implementations of the reinforcement learning-based experimental teaching effectiveness evaluation device of the present invention can be referred to the above-described method embodiments, and will not be repeated here.

[0127] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A method for evaluating the effectiveness of experimental teaching based on reinforcement learning, characterized in that, The method includes: A number of experimental courses are set up in the teaching system. Multiple influencing factors that affect the teaching effect of the experimental courses are obtained. Based on the influencing factors, the experimental courses are classified to obtain multiple course groups consisting of at least one experimental course. The influencing factors include teaching environment parameters, student characteristic parameters, and experimental characteristic parameters. Determine the preset evaluation parameters for assessing teaching effectiveness, and assign different parameter weight values ​​to the evaluation parameters according to the analytic hierarchy process. For the course group, a preliminary evaluation of the teaching effectiveness is conducted based on the parameter weight values, resulting in a group evaluation result composed of individual evaluation results corresponding to multiple experimental courses. Based on the group evaluation results, target evaluation parameters for evaluating the teaching effectiveness of the corresponding course groups are selected from the evaluation parameters; For each experimental course, a membership function corresponding to the target evaluation parameter is constructed, and the target evaluation parameter value corresponding to the experimental course is collected. Based on the membership function and the target evaluation parameter value, the membership degree of the target evaluation parameter value within the preset evaluation level is calculated. Based on the parameter weights and membership degrees, the teaching effectiveness of the experimental course is dynamically optimized and evaluated by a reinforcement learning agent to determine the final teaching effect of the experimental course.

2. The method for evaluating the effectiveness of experimental teaching based on reinforcement learning as described in claim 1, characterized in that, The process of classifying the experimental courses according to the influencing factors yields multiple course groups, each consisting of at least one experimental course, specifically including: Based on the aforementioned impact factors, determine the impact factor values ​​corresponding to each of the experimental courses; From the aforementioned experimental courses, select designated experimental courses whose differences between the values ​​of each influencing factor are within a preset range to form a course group, thereby classifying the aforementioned experimental courses. The preset range is dynamically adjusted according to the reinforcement learning training stage. A more lenient threshold is used in the initial training stage, and a more stringent threshold is used in the convergence stage.

3. The method for evaluating the effectiveness of experimental teaching based on reinforcement learning as described in claim 2, characterized in that, Before selecting a specific experimental course from the plurality of experimental courses, the method further includes: Based on the historical performance data of the aforementioned experimental courses, a corresponding learning behavior change curve is generated, and based on the learning behavior change curve, the teaching stage of the current experimental course is estimated through the reinforcement learning state recognition module. The range of differences between the influencing factor values ​​is dynamically adjusted according to the teaching stage; wherein, the range of differences is positively correlated with the students' adaptability at the teaching stage: the lower the adaptability, the wider the range of differences.

4. The method for evaluating the effectiveness of experimental teaching based on reinforcement learning as described in claim 1, characterized in that, The step of selecting target evaluation parameters for evaluating the teaching effectiveness of the course group from the evaluation parameters based on the group evaluation results specifically includes: The fluctuation value between individual evaluation results corresponding to any two experimental courses in the group evaluation results is determined by a reinforcement learning policy network, and the maximum fluctuation value is selected from the fluctuation values. If the maximum fluctuation value is within a preset range, the evaluation parameter is used as the target evaluation parameter; Otherwise, the reinforcement learning action selection module is invoked to combine the evaluation parameters to generate an evaluation parameter set. Based on the evaluation parameter set, the evaluation strategy is optimized through multiple rounds of iterative training until the fluctuation values ​​corresponding to the parameter set all meet the preset range. Then, the evaluation parameter set is used as the target evaluation parameter.

5. The method for evaluating the effectiveness of experimental teaching based on reinforcement learning as described in claim 4, characterized in that, After iterative evaluation via reinforcement learning, the method further includes: When multiple evaluation parameter sets have fluctuation values ​​that meet the preset range, calculate the cumulative reward value for each parameter set; Based on the cumulative reward value, the set of evaluation parameters with the highest reward value is selected as the target evaluation parameter; wherein, the cumulative reward value is calculated by the reinforcement learning agent based on the improvement of teaching effect and evaluation stability.

6. The method for evaluating the effectiveness of experimental teaching based on reinforcement learning as described in claim 1, characterized in that, The construction of the membership function corresponding to the target evaluation parameters specifically includes: The evaluation levels corresponding to the target evaluation parameters are divided; wherein each evaluation level is mapped to a state space interval in reinforcement learning. For each target evaluation parameter, the Q-learning algorithm is used to optimize the level membership degree of the parameter in each level interval. The level membership degree is then re-standardized based on reward feedback to obtain the membership degree function corresponding to the target evaluation parameter.

7. The method for evaluating the effectiveness of experimental teaching based on reinforcement learning as described in claim 6, characterized in that, Determining the final teaching effect corresponding to the experimental course specifically includes: An evaluation vector is generated based on the membership degree of the target evaluation parameter value within each evaluation level; The evaluation vectors of all target evaluation parameters are concatenated into a state feature matrix, which is then input into the reinforcement learning evaluation model. The reinforcement learning evaluation model performs fuzzy transformation and reward update through its policy network, and outputs a confidence vector for teaching effectiveness. Based on the optimal action selected by the reinforcement learning agent in the teaching effect confidence vector, the final teaching effect corresponding to the experimental course is determined.

8. The method for evaluating the effectiveness of experimental teaching based on reinforcement learning as described in claim 7, characterized in that, The determination of the final teaching effect of the experimental course based on the optimal action selected by the reinforcement learning agent in the teaching effect confidence vector specifically includes: The evaluation level with the highest probability value selected by the reinforcement learning agent is selected from the confidence vector of the teaching effect. If the selection probability value is lower than a preset threshold, the exploration mechanism is triggered to regenerate the evaluation parameter set and return to the iterative evaluation step; Otherwise, the assessment level will be taken as the final teaching effect, and the assessment results will be stored in the experience playback pool to optimize subsequent strategies.

9. A device for evaluating the effectiveness of experimental teaching based on reinforcement learning, characterized in that, The device includes: a memory, a processor, and a reinforcement learning-based experimental teaching effectiveness evaluation program stored in the memory and executable on the processor, wherein the reinforcement learning-based experimental teaching effectiveness evaluation program is configured to implement the steps of the reinforcement learning-based experimental teaching effectiveness evaluation method as described in any one of claims 1 to 8.

10. A medium, characterized in that, The medium stores a reinforcement learning-based experimental teaching effectiveness evaluation program, which, when executed by a processor, implements the steps of the reinforcement learning-based experimental teaching effectiveness evaluation method as described in any one of claims 1 to 8.

Images