Dimensionality reduction of RKHS model parameters.

Area of Science:

• Machine Learning
• System Identification
• Nonlinear Dynamics

Background:

• Models in Reproducing Kernel Hilbert Space (RKHS) often require a high number of parameters, correlating with the learning phase's observations.
• Efficient model reduction is crucial for practical applications in complex systems.

Purpose of the Study:

• To introduce a novel method, Reduced Kernel Partial Least Square (RKPLS), for decreasing the parameter count in RKHS models.
• To compare the performance of RKPLS against the Support Vector Machines on Regression (SVR) technique.

Main Methods:

• The RKPLS method approximates latent components from Kernel Partial Least Square (KPLS) using nearest observation vectors.
• A comparative study was designed to evaluate RKPLS against SVR.
• The methods were applied to identify a nonlinear Process Trainer PT326, a laboratory physical thermal process.

Main Results:

• The proposed RKPLS method demonstrated satisfactory performance in identifying the nonlinear Process Trainer PT326.
• Results indicate that RKPLS is a viable alternative to SVR for this type of identification task.

Conclusions:

• RKPLS offers an effective strategy for reducing model complexity in RKHS.
• The method shows promise for identifying nonlinear dynamic systems, particularly those with thermal characteristics and large time responses.