Research on Optimising Personalised Teaching Models in University Piano Courses Using Reinforcement Learning Algorithms

With the rapid development and popularisation of information technology, the field of education is constantly changing. The core of traditional education is still the teacher occupies a dominant position, the students are in passive acceptance, and the effect of student learning depends to a certain extent on the student's self-driven force, which to a certain extent restricts the comprehensive advancement of education. Passive education model is not only reflected in mathematics, physics and other traditional subjects, but due to the state, society, schools and families to pay attention to and guidance, in the traditional learning process can ensure sufficient time and training to ensure the learning effect, where the active learning and passive learning, learning interest and the importance of personalised teaching mode is not fully reflected. [1] In the case of arts, music subjects have not been given the same level of attention and guidance as traditional subjects in the traditional education model, which has led to the failure of mandatory practice and time guarantee for learning in the society, school and even at home. Failure to get mandatory practice and time guarantee, which to a certain extent inhibits the development of students' individuality. [2] However, with the rapid development of economy, the social requirements for However, with the rapid development of the economy, the requirements of the society for students are becoming more and more diversified and comprehensive, which in turn requires the society, schools and families to pay attention to the personalised development of students, but the degree of such attention is still weaker than that of traditional disciplines. [3] We conducted random interviews with students and parents as shown in Figure 1. 1 shows, we found that roughly 71% of people think that music class is not valued, and only 24% think that it is valued, indicating that the music subject is not recognised by the majority of people, which suggests that the music teaching mode needs to be further improved and optimised. The development of the art and music disciplines therefore remains largely dependent on the students' own personalised approach to learning, which relies on the ability to analyse their own situation and designate a personalised plan, often relying on a personalised teaching model.

The integration of technology has given rise to a new development path: personalised learning empowered by technology. [4] Especially after entering the information age, the large-scale popularisation and application of Internet technology and artificial intelligence technology have begun to significantly reflect the theme of technology changing life. Especially in the field of education, the rapid development of information technology has provided new possibilities for distance education and blended learning models. Blended learning environments combine the advantages of face-to-face and virtual teaching and aim to provide learners with a more flexible and personalised learning experience. The emergence of personalised adaptive learning is due to the rise of big data technologies, where data is being generated in more and more ways and at a faster rate, which has given rise to a fourth paradigm of scientific research - data-intensive science. [5] Under the influence of data-intensive science, personalised adaptive learning has become a hotspot in current educational technology research.

Current technological developments have made personalised learning more and more adaptive and adaptive learning more and more personalised. This trend is even more evident in learning environments that support smart technologies. [6] At the beginning of the learning algorithm model used in personalised teaching is Skinner's teaching machine and program teaching theory for the first time embodied, but limited by the conditions of computing services at the time did not apply on a large scale. And with the development of technology, personalised learning has gradually become more complex. Typically represented by the KNOWLEDGE platform, which differentiates and guides the ongoing process of personalised learning through knowledge maps, real-time monitoring and responding to student activities (communication, collaboration and games). The new personalised learning approach is proactive and makes decisions based on data collected by automated systems; it adapts to the real-time learning conditions of the learner so that the content and activities meet the learner's individual characteristics and needs. [7] Automated systems collect and analyse comprehensive information, which was unimaginable in the past, but with the rise of the big data era, this makes it possible to record and interpret students' personal characteristics and real-time status in all aspects of learning. And personalised adaptive learning, which will accompany this trend for a long time, will be the main development model of education in the future. [8] And for learning personalised learning style, it needs school teachers to develop corresponding personalised teaching mode. As an art music discipline, piano class in university may be the first time for most students to receive a more systematic music training, and the learning of this course precisely emphasises the cultivation of practical and application ability. However, traditional university music classroom teaching is often faced with the problems of monotonous content, slow teacher-student feedback and low student participation, which further limits the teaching effect and students' personalised development. In an information-based social environment, the piano teaching model is shifting from the traditional teacher-centred approach to a more interactive, multimedia and personalised direction.

Therefore, based on the above background and motivation, this study aims to conduct an empirical study to gain a deeper understanding of the application of machine learning in personalised teaching models in university piano teaching. The aim of this paper is to analyse in-depth the piano classroom teaching model in an information-based environment and to effectively explore its potential to improve teaching efficiency, student motivation and teaching quality. Therefore, this paper explores how deep learning technologies can be used to improve university piano classroom teaching models, especially in terms of personalised learning paths and resource recommendations. This study aims to explore the impact of digital reforms on improving teaching efficiency and quality, thereby providing valuable insights and recommendations for piano teaching practices and educational technology research in universities.

2

Learning algorithm model of personalised teaching mode

2.1

Analysis of the existing model of the university piano classroom model

Intuitively, the personalised teaching model focuses on information about individual differences in order to develop relevant teaching tasks. Instead of the traditional one-size-fits-all approach, teachers should rationalise students' lesson plans according to their individual needs, the gap between the current state and the learner's expected state. The differences in students' needs are determined by the current state and status quo, and in today's information technology this information can be obtained and analysed through relevant technology. According to the Association for Supervision and Curriculum Development (ASCD) related definition of individualised instruction: differentiated instruction is a type of instruction in which educators actively plan for student differences so that all students learn best. In a differentiated instruction classroom, teachers allocate their time, resources, and energy to effectively teach students with different backgrounds, preparation, skill levels, and interests. [9]

In the current university piano teaching there are various teaching modes, including traditional teaching mode, interactive teaching mode and flipped classroom teaching mode, and there are also specific personalised teaching modes. The traditional teaching mode is usually teacher-centred, with the teacher mainly delivering the content and the students passively accepting the knowledge, resulting in a relatively monotonous learning style for the students. Coupled with the fact that interactive classrooms are difficult to interact with, it is difficult for teachers to grasp the basic ability and learning of students, leading to interactive classrooms exist in name only. The flipped classroom is more demanding than the former classroom environment, which can strengthen the students ‘participation and enthusiasm in the piano classroom, and facilitate the students to show their personalities and interests, but due to the difficulty of taking care of every student in the flipped classroom, coupled with the students’ herd and blind obedience, it is difficult for them to express their personal reality, so they cannot formulate a personalised teaching plan for individuals clearly and accurately. [10] And personalised teaching mode requires information technology, high degree of intelligence, and relatively high teaching costs, but the teaching mode can quickly analyse the students' personal situation and formulate the corresponding personalised teaching mode. This teaching mode can actively mobilise students' enthusiasm for learning, and at the same time, it can avoid the above mentioned mentality of blindly following the herd, and it can achieve timely information feedback in the classroom. We collected the satisfaction degree of 1000 college students in a city on the college piano course to reflect the impact of the teaching mode on the students. We can find from Table 1 that the satisfaction rate of traditional model is 68.3%, interactive model is 73.6%, flipped classroom model is 81.6%, and the personalised teaching model with the highest satisfaction rate is 94.7%. With the increase of classroom interactivity, participation and intelligence, the satisfaction rate of students gradually increases, which proves that the personalised teaching model well meets the needs of contemporary university students in piano course learning, and also proves the importance of machine learning optimisation for personalised teaching model. We will consider in the following work how personalised recommender systems can be integrated into these models and how the model can be optimised by machine learning to support distance education and blended learning environments.

Table 1.

Satisfaction with different music classroom models among 1000 A-city university students

	Traditional teaching model	Interactive teaching mode	Reverse teaching model	Personalized teaching model
Satisfied	683	736	816	947
Unsatisfied	317	264	184	53
Satisfaction rate	68.3%	73.6%	81.6%	94.7%

2.2

Analysis of the construction of personalised teaching for piano teaching based on deep learning algorithms

The machine optimisation of the personalised system model is based on personalised information about students' learning interests, levels and learning history to personalise the teaching model in order to improve learning efficiency and interest. As shown in Table 2, it demonstrates the current distribution of teachers and students involved in personalised music education in university teaching. We can know that the degree of participation in personalised music teaching by both student and teacher groups is still relatively low at present, which also provides a practical need for us to establish a model to optimise the personalised teaching mode. At the same time, in order to achieve the optimisation of personalised content for university piano teaching, we use the automatically collected information mentioned above as the main information input, use machine learning algorithms to identify features in the text, use a graph convolutional neural network (GCNN) model to extract the features of the personalised teaching quality information from the images or videos, and then use an attention mechanism to fuse and analyse these features. From this we construct a model consisting of an information collection layer, a graph convolutional network layer, an attention layer and a classification layer. [11] Different personalised teaching programmes are developed based on the feedback of different students ‘models.In the process of developing the teaching model, according to the different information of learners’ learning status, level, etc., it may be necessary for the system to adopt different matching strategies to establish different lists of learning units, and learners can select the most suitable learning units according to their own needs. Learners can choose the most suitable list of learning units according to their own needs. The system then measures the learning effect and generates the learner's learning profile. This is an iterative process of generation and optimisation; it allows the learner to have the autonomy of learning rather than being forced to recommend learning content, which highlights the learner's initiative and subjectivity, and also meets the requirements of the personalised teaching model of the university piano classroom.

Table 2.

Statistics on the participation of different teachers and students in personalised music teaching

Project	Number of teachers	Proportion	Number of children	Proportion
Never used	46	46%	338	68%
Occasionally use	38	38%	107	21%
Frequently use	16	16%	55	11%
Total	100		500

The optimisation model is shown in Figure 2, in which the data from the teaching content of the university piano class was first preprocessed. Then, the information gathering layer and graph convolution layer are used for student feedback feature extraction. The information gathering layer uses a vector matrix to convert the input text into a real-valued vector representation. The signal vector Ea for a single grouping can be represented as (1) $E_{a} = a_{t} w_{t}$ (2) $w_{t} \in R^{| A | e}$

The wt representation is the vector matrix of the training signal, A denotes the magnitude of the signal strength, and e denotes the vector dimension.

Traditional differentiated instruction strategies are divided into subgroups: homogeneous subgroups and heterogeneous subgroups. But that stratification is solidified and roughly differentiated as a representative form of instruction. Dynamic grouping and intervention has become possible in the IT environment. We therefore use a dynamic grouping model in our model, where different learning unit group vectors are typically used to indicate to which signal zone the signal belongs to, in order to facilitate subsequent coding processes. This vector group can be represented by Eb using the equation (3) $E_{b} = a_{b} w_{b}$ (4) $w_{b} \in R^{| B | e}$

Eb denotes the vector group encoding is converted to a real-valued vector by this vector matrix wb, wb denotes this vector matrix operation, and |B| denotes the number of cell groups.

And for the determination of the position of different signal vectors, we need position vectors for encoding the absolute position of each word, which can be expressed as the equation (5) $E_{c} = a_{c} w_{c}$ (6) $w_{b} \in R^{| N | e}$ w_c denotes the position vector and N denotes the maximum position length.

Reinforcement learning algorithms to optimise information feedback design

The information collection layer model has three important characteristics: accuracy, personalisation and generality. [12] In order to improve the information representation and processing ability of the information gathering and graph convolutional neural network model, we train a personalised teaching model for university piano lessons by masking the signal model task. In this structural layer various comprehensive information is transformed into machine language and visual graphs through an automated information collection system, in the data collection and processing, which is the first step to optimise the model. [13] The accuracy of information gathering is the first step that directly affects the success or failure of machine learning optimisation. In this process we train the model's accuracy on classroom effects through an explicit or semi-explicit source of signals, such as actions, expressions, and other tasks, so as to better support downstream analyses and other session tasks.

Definition for graphical convolutional neural network models (7) $G = (X, Y)$ (8) $e_{ij} \in (X_{i}, X_{j}) \in Y$ (9) $X_{i}, X_{j} \in X$ (10) $d_{i} = concat (K_{ab (x_{i})}, K_{ab (x_{j})})$

The set of information is defined as G, where eij is the set of boundaries and the edge weights wij, where di denotes the feature vector of the first vertex Xi. The above processed and transformed information data is used as input data for GCNM.

And the calculation method of GCNN is shown in the following equation (11) $H^{(1 + L)} = f ({\bar{D}}^{- 1 / 2} \bar{A} {\bar{D}}^{- 1 / 2} H^{L} Q^{L})$ (12) $\bar{D} = \sum_{j} {\bar{A}}_{i j}$

H^L is the hidden layer in the first layer, which belongs to the first input signal, $\bar{A}$ denotes the adjacency matrix of the main interaction graph, and $\bar{D}$ denotes the degree matrix containing the coefficients of each vertex in the interaction graph. Q^L is the parameter matrix learned by the network, and f is the activation function, which employs the activation of ReLU. Finally, the vector representations of all the vertices in the terminal hidden layer must be centralised into a single vector and denoted as Kab. the vector Kab is then fed into a multilayer computational parameter in order to compute the final matching score.

And in the attention layer, we utilise a multi-head attention graph network model to enhance the accuracy of the model. We optimise the original model for fast signal capture and processing. The output of connecting multiple sensors can be expressed as the equation (13) $O_{v} = \sum_{t=1}^{T} a(\sum_{u \in N_{v}} {wa}_{vu} h_{u})$

O_v is the value of the signal input strength, T is the number of heads of note, and Nv is the first-order neighbourhood of vertex v. Nv denotes the normalized correlation coefficient between the vertex v of the tth head and its neighbouring vertices.

Furthermore, the result of averaging multiple heads can be expressed by the equation (14) $o_{av, v} = \frac{1}{T} \sum_{t=1}^{T} a(\sum_{u \in N_{v}} {wa}_{vu} h_{u})$

O_av,v is the average of the signal input strength.

The graph attention mechanism layer is used to homogenise the local feature vectors with the matching feature vectors and gives a new method of calculating the feature vector KAB, which is processed by the following equation (15) $K_{ab} = GAT (K_{ab (x_{i})}, ‖ {GATK}_{ab (x_{j})})$

K_{ab(X_i)} and K_{ab(X_j)} represent the feature vectors and matching feature vectors in the mechanism layer, respectively.

Finally, the prediction layer predicts the final score for the personalisation effect of the piano teaching content based on the vector calculation process described above. The final output vector OUT can be expressed as the equation (16) $OUT =$

We have designed a dynamic optimisation model taking into account vectors that are constantly changing and optimised for individual groupings during machine learning. In the process of optimisation of this model, it is necessary to constantly optimise the regulation of the grouping of students in order to target the requirements of personalised teaching in the course in order to obtain better personalised teaching in order to improve the quality of teaching. The process of regulating groups is shown in Figure 3. The system adopts different grouping matching strategies to recommend a suitable group according to the learner's learning status for their own personalised teaching, and the learner can choose the most suitable learning programme according to their own needs. The system then measures the learning effect and then generates the learner's learning profile in order to continue optimising the grouping. The model is better able to accurately develop a suitable personalised teaching model for each student, as well as quickly grouping students according to their individual profiles, which relies on the accuracy of the model and the selectivity of the dynamic optimisation.

Therefore theoretically we design a personalised optimisation system model for university piano courses to better meet students ‘individual learning needs, improve learning efficiency and effectiveness, and enhance students’ humanistic level, as well as to promote the innovation and development of university piano teaching content and methods. Specifically feedback information is obtained in favour of the curriculum level including institutional changes, such as the elective course system, elective course system, credit system and so on. The curriculum level includes the planning of schools and teachers for students ‘learning paths, while the task level involves the adaptation and regulation of teachers’ specific teaching content, services, etc.

2.3

Determination and optimisation of machine learning process parameter weights

With the progress of educational technology and the rise of the concept of personalised education, schools and students are increasingly demanding for personalised teaching models, and the continuous intelligence of personalised teaching models, digital has subverted the traditional teaching model. [14] The recommendation design system for personalised teaching mode is based on deep learning techniques to analyse and extract multimedia features in teaching materials, such as images and video content. Through the automatic information collection system and attention mechanism, the system can achieve the integration of feature signals from different sources, so as to generate comprehensive student personalised information and provide teachers with corresponding targeted recommendations, thus providing personalised learning modes and solutions for university piano classroom teaching. When the system analyses the signal data, it needs to design different signal groupings and weightings according to the actual situation, which can be used as an effective feedback for judging the quality of students' teaching in the classroom. We designed the distribution table of the percentage of different signal sources according to the original intention of the university piano course design, as shown in Table 3. Since the main purpose of the university piano course is to improve college students' music appreciation ability and systematic professional training, in order to achieve the purpose of selecting relevant professionals. Therefore, in the weighting table, we greatly weakened the index of talent, strengthened the students ‘own interest and personal level and students’ own personalised performance and cultivation of parameter weighting, which is conducive to more effective differentiation of the real situation of the students, to achieve the goal of each student's ability to improve music appreciation. At the same time, because college students are in a more complex learning and social environment compared to the previous learning environment, which leads to student feedback and information processing is prone to recurrence, in order to better improve the accuracy of the model, we added a consistency indicator, the indicator is determined after each weight according to the student's performance in the collection of information many times, to avoid the occurrence of chance and recurrence affect the model of the information error.

Table 3.

Allocation of weights for student performance in university music teaching classroom

Project	Proportion	Duration	Proportion
Level of interest	20%	Never used	10%
Classroom performance	20%	Occasionally use	30%
Talent	10%	Frequently use	60%
Level of diligence	20%
Classroom feedback situation	30%

The quick calculation method is: (17) $\sum_{i=1}^{n} project * duration$

3

Experimental results and discussion

In order to improve the teaching quality of university music education courses and design a personalised teaching model, we improve the deficiencies of the existing traditional education courses through the optimisation of reinforcement learning algorithms. 500 university students in M city were selected for the survey, and the characteristics of each sample are the corresponding evaluation indexes, and to ensure that teachers can get timely feedback from the corresponding model and adjust the course in time. This work uses accuracy and satisfaction rate to evaluate the actual network performance. In this computational method we establish a set of evaluation indicators that enable the level of quality of teaching in university music courses, which involves the use of information automatically collected by artificial intelligence, such as students' classroom performance, the degree of practice, etc. At the same time, we downplay factors such as natural talent, and make clear the objectives of university music teaching, so as to establish an optimised model for the allocation of the weights of the evaluation system. As shown in Table 4, we demonstrate the design process of the teaching experiment of the model. The results of the model have only two outcomes, satisfactory and unsatisfactory. We used the above programming model system as the calculation and calibration method and the above indicators as the evaluation parameters to statistically analyse the data of 500 children.

Table 4.

Design of the teaching experiment

	Control group	Experimental group
Students	500 students in c city	500 students in c city
Enviroment	c city	c city
Expeiment	Traditional teaching mode	Personalized teaching mode
Variables	Reinforcement Learning Algorithm Optimization System

Prior to the experiment, we needed to statistically analyse the experimental group of students against the control group of students to ensure the relative agreement of the experimental samples. We took two groups of students, the control group and the experimental group, in a course teaching mode similar to music, to analyse their differential situation to ensure that the differences between the two comparison groups were minimised, and the results of the differential statistics are shown in Figure 4. In this experiment the experimental group and the control group of students performed similarly in terms of their independent learning ability, participation, problem solving ability and efficiency in the use of classroom time in the university music classroom, with no statistical difference. The proportion of consistency between the experimental group and the control group on each dimension at the end of the experiment was further analysed.

By using the above optimisation model to adjust and improve the teaching effect of a music course of 16 lessons and 32 hours in one semester, we found that the personalised teaching mode of constantly adjusting the course design and strengthening the students' grouping in time during the teaching process after using the optimisation model is more likely to require good feedback and teaching effect than the traditional teaching mode, as shown in Figure 5. We found that in the traditional classroom teaching model without the optimisation model, after one semester of study, the number of students who rated their teaching as satisfactory increased from 300 to 400, and remained relatively stable, but there was a noticeable repetition of the number of students throughout the teaching process. In the model-optimised personalised teaching model, under relatively consistent conditions, the number of students achieving satisfactory teaching ratings increased from 300 to nearly 500, and reached a relatively stable number of students more quickly, and the rate of growth of students achieving the teaching goals was significantly faster than that of the traditional teaching model that was not optimised with the relevant model. The optimised personalised teaching model achieves better results in a shorter period of time than the traditional teaching model, and the personalised curriculum is designed to increase the number of students at the upper limit of the quality of teaching, while avoiding to a large extent the repetitive performance of students in the teaching process. This depends largely on the model's timely feedback and accurate grouping and evaluation system.

Figure 6 shows the changes in university students' satisfaction with different teaching models over the course of a semester, and we find that the satisfaction rate of students in the personalised teaching model is consistently higher than the satisfaction rate of teaching in the traditional model. In the 10th course, the satisfaction rate of students in the personalised teaching mode is close to 100%, while the upper limit of the satisfaction rate of the teaching mode of the traditional model is only close to 80%, and the upward trend of the satisfaction rate is much smaller than that of the optimised personalised teaching mode, which demonstrates that the personalised teaching mode has a better relevance and satisfaction rate relative to the traditional teaching mode, which is in line with the conclusions obtained above. Figure 7 shows the distribution of college students' dissatisfaction with different music classroom evaluations, and we can find that the number of college students who are dissatisfied with the personalised teaching model of the optimised model is significantly lower than the number of dissatisfied with the traditional teaching model during the teaching cycle. After a semester of personalised learning, almost none of the 500 people gave the teaching model dissatisfaction, while the number of dissatisfied people in traditional teaching was almost 100, this part of the students thought that the classroom did not give them what they needed for that music classroom, which may be because this part of the students may be more demanding, or they themselves are not good at information feedback with the teacher, which limits the sustainable improvement of the quality of the teaching and learning, which is also is one of the common disadvantages of the traditional teaching model. At the same time, the personalised teaching model has a lower level of dissatisfaction, which is due to the fact that the personalised teaching model can design relevant courses according to the needs of different groups of students in university music classes. At the same time, in order to address the situation of a large number of students in university teaching and the complexity of the situation, the optimisation model adopts the dynamic grouping model to realise optimal grouping of different groups of students with different needs, and to set learning plans at regular intervals.

4

Conclusion

We have established a set of dynamically grouped weights combined with a practical personalised teaching optimisation model through an in-depth analysis of the patterns and educational goals that exist in university music classroom teaching, combined with a learning algorithm model. The model system combines the personalised teaching mode and the blended learning environment, which can be found to have an obvious positive impact on the students' learning effect and teaching effect. By providing customised dynamic grouping with personalised learning plans, these are designed to meet the individual needs of students, thereby stimulating their interest and motivation in learning. The results of the experiment showed that students in the experimental group showed significant improvements in their independent learning ability, learning satisfaction and engagement, which suggests that the Personalised Teaching Optimisation model system can promote active learning and deep engagement. In addition, the personalised teaching optimisation system can also improve teaching effectiveness. Teachers can better understand students' learning based on the data and feedback provided by the system in order to adjust teaching strategies and content for more targeted teaching. This data-driven teaching approach helps improve the quality and efficiency of teaching. At the same time, the more objective evaluation tools and techniques we have established, such as dynamic data analysis and rapid grouping system, can improve the accuracy of evaluation and reduce computing time. For the design of other personalised education models, other education researchers can further optimise the existing models, explore their applications in different teaching disciplines, and investigate how to integrate them with existing educational technology tools and platforms. In addition, this paper has not specifically analysed and modelled students' private learning processes, family factors, etc. Therefore studying the ethical and privacy issues that should be incorporated in personalised teaching systems will be an important area for future research. Through these efforts, it is expected to promote innovation in educational technology and provide learners with a more personalised and effective educational experience. Future research will further explore the use of personalised recommender systems in a wider range of personalised education systems and blended learning environments.

Sprache:: Englisch

Zeitrahmen der Veröffentlichung:: 1 Hefte pro Jahr
Fachgebiete der Zeitschrift:: Biologie, Biologie, andere, Mathematik, Angewandte Mathematik, Mathematik, Allgemeines, Physik, Physik, andere

Zeitschrift RSS Feed

Research on Optimising Personalised Teaching Models in University Piano Courses Using Reinforcement Learning Algorithms

Xiaoxuan Zhang

Online veröffentlicht: 27. Feb. 2025

Eingereicht: 23. Okt. 2024

Akzeptiert: 26. Jan. 2025

DOI: https://doi.org/10.2478/amns-2025-0110

Schlüsselwörtermachine learning, personalised teaching, dynamic self-regulation, university education

© 2025 Xiaoxuan Zhang, published by Sciendo

This work is licensed under the Creative Commons Attribution 4.0 International License.

Schlüsselwörter
machine learning, personalised teaching, dynamic self-regulation, university education