Abstract
In order to address the shortcomings of traditional teaching quality evaluation systems, researchers have implemented Particle Swarm Optimization (PSO) neural network technology to develop a teaching quality evaluation system for colleges and universities. This system utilizes PSO neural network technology, with teaching evaluation standards as inputs and teaching evaluation results as outputs. By using Anhui University of Science and Technology as a case study, empirical research demonstrates that this approach not only mitigates the direct impact of human factors on evaluation outcomes but also provides a basis for establishing a comprehensive and rational evaluation index system.
Introduction
Evaluating the quality of higher education is notoriously difficult. Traditional systems often rely on subjective surveys or rigid percentage-based metrics that fail to capture the nuance of effective teaching. The relationship between "teaching inputs" (such as a professor's enthusiasm or preparation) and "outputs" (such as student analytical ability) is highly nonlinear and influenced by numerous qualitative factors.
Furthermore, standard evaluation methods are often prone to "subjective arbitrariness"—human bias that can skew results. The authors aim to create a more scientific, automated, and objective framework that can process these complex variables to provide actionable feedback for educators.
Methodology
The core of this research is the PSO-BP Neural Network, which combines the predictive power of neural networks with the global search capabilities of swarm intelligence.
Backpropagation (BP) Neural Network
The authors utilize a standard three-layer architecture:
- Input Layer: 15 neurons representing specific evaluation criteria (e.g., teaching attitude, clarity of explanation, use of frontier knowledge).
- Hidden Layer: 8 neurons, determined by convergence performance.
- Output Layer: 1 neuron representing the final evaluation score.
Particle Swarm Optimization (PSO)
Standard BP networks often struggle with "local optima"—getting stuck in a sub-optimal solution during training—and slow convergence. PSO is an evolutionary computing technique inspired by the social behavior of bird flocking. In this model, "particles" (potential solutions) fly through the solution space to find the optimal weights and biases for the neural network. By using PSO to initialize and optimize the neural network's parameters, the system achieves better global search capability and faster computing speeds than a standalone neural network.
Empirical Research and Datasets
The paper conducts empirical research using a case study from the Anhui University of Science and Technology.
- Input Data: The model processes 15 distinct evaluation criteria ($X_1$ through $X_{15}$), ranging from "teaching enthusiasm" and "vividness of language" to "cultivation of innovative consciousness" and "improvement of problem-solving abilities."
- Quantification: Rather than a simple percentage, the system uses an A-E grading scale, which is then converted into quantitative data for the neural network to process. This allows the system to simulate human judgment while maintaining algorithmic consistency.
Key Insights and Results
The most significant insight is the application of PSO to overcome the inherent weaknesses of BP neural networks. While gradient descent (standard BP) can get "lost" in complex data landscapes, the "flocking" nature of PSO ensures the model explores the entire solution space for the best possible configuration.
Key findings include:
- Mitigation of Human Factors: The empirical results demonstrate that the PSO-NN approach significantly reduces the direct impact of human subjectivity on evaluation outcomes.
- Reliability and Speed: The PSO-BP model overcomes slow convergence and provides more scientific, "rational" results compared to traditional mainstream evaluation systems.
- Automated Feedback: The system identifies specific indicators with low scores, allowing universities to provide targeted feedback to teachers to improve classroom performance.
Conclusion and Future Work
The researchers highlight that the entire process—from index quantification to data processing—can be fully realized through computer programming, making it a highly scalable solution for large institutions.
Future work suggested by the study includes further refining the mathematical models to handle even more "non-quantitative" factors and expanding the index system. Additionally, testing the model's generalizability across different academic disciplines—such as comparing lab-based science courses versus theoretical humanities seminars—remains a key area for further exploration.
Summary updated