Benchmarking of the Key Sample Machine

Liu, Yu (2005) Benchmarking of the Key Sample Machine.

[img]
Preview
PDF
888kB
Abstract:In learning control schemes such as Learning Feed-Forward Control(LFFC), a function approximator is embedded to approximate the inverse of the plant under control. The function approximation concerns itself with finding a relation within a set of samples supplied before. The goal of the function approximator is to select the best function from a set of possible functions to approximate the true relation contained in the noisy data set. In previously LFFC research, different function approximator has been used: the Least Squares Approximator, the neural networks such as Multilayer Perceptron (MLP)and B-splines Networks, the Support Vector Machines (SVM) and the Least Squares Support Vector Machines (LSSVM), ect. Recently a new off-line approximator called Key Sample Machine (KSM) has been developed during the Ph.D research of dr. ir. B.J. de Kruif. A series of tests had been made to evaluate the performance of KSM and showed it gives a smaller prediction error for noisy data in a shorter time compared with other methods. However, these tests have not been performed with a sufficient number of different problems/data sets. With a smaller volume of data sets, it is impossible to characterize the behaviour of KSM in a broad statistical view. For this purpose a standard of experimentation should be valid in the sense that there are no artefacts created by random factors or by a faulty experimental setup. The goal of this research is to test and evaluate the KSM algorithm by standard experiments and to compare it with other available methods (SVM,LSSVM,RBFN) on standard sets of regression problems through simulation. Nine well-known approximation benchmarking problems are used in our simulation, including six real-world data sets and three artificial data sets. The comparison is performed with respect to the ability to approximate various data sets, the efficiency of learning, and the ability to train samples from the real-world data. Simulation results show that the KSM yields good performance, especially when considering the efficiency, which is very important for its application in the online LFFC setting. In short, our results confirm the potential of KSM to be a good function approximator.
Item Type:Essay (Master)
Faculty:EEMCS: Electrical Engineering, Mathematics and Computer Science
Subject:53 electrotechnology
Programme:Electrical Engineering MSc (60353)
Link to this item:http://purl.utwente.nl/essays/56917
Export this item as:BibTeX
EndNote
HTML Citation
Reference Manager

 

Repository Staff Only: item control page