Function Approximation through an Efficient Neural Networks Method

Abstract

Abstract Neural network system, a portion of artificial intelligence, is increasingly becoming prevalent nowadays. The application of neural networks models to function approximation is one of the latest developments in electrical engineering including robotics route planning. It is, however, a challenging task to instruct on this topic in artificial intelligence courses. In general, a function approximation issue aims to select a function among a well-defined class that closely matches a target function in a task-specific manner, which has a large number of applications in engineering such as the robotics route planning. In this paper, we present how we follow the pedagogies of sparrow-dissection and scaffolding and help students design, implement, debug, and operate an efficient neural networks model method for function approximation. A back propagation neural networks model is taught in the classroom with the source code provided to students. The students are required to revise and modify the source code therefore apply the back propagation neural networks model for the function approximation. Specifically, in our Artificial Intelligence Techniques course for senior and graduate students, we provide a back propagation neural networks model with its source code as a ‘sparrow’ for students to dissect and examine how the neural networks model for function approximation works. We then work together with students in revision and modification of the source code for purpose of function approximation. With the gradual withdrawal of the instructor in a series of projects assigned to students on the application of back propagation neural networks model to function approximation in robot path planning, students gain more independence and can complete the projects fully by themselves in the end. A set of projects are assigned to students to perform the neural networks model for the function approximation. An assessment approach is given to determine how comfortable students are with neural network concepts before and after the project. In addition, feedback was solicited after each neural network models to get feedback from students about implementation, development and application of the models into function approximation. The effectiveness of the neural network model for function approximation is evaluated through various milestone assignments, lab reports, presentations, and other activities. In addition, we survey students for their feedback on how comfortable they are with neural network concepts before and after each project, as well as feedback on the process of development and application of each neural networks model by revising the first source code of neural network model. Teaching and learning strategies by the neural network model methodology were associated with learning outcomes of this course by analysis of the function approximation project. All these data, together with from those from the official course evaluation system, point to the effectiveness of the neural network model approach and high learning quality as a result of our sparrow-dissection and scaffolding pedagogies.