Probabilistic characterization and synthesis of complex driven systems

Abstract

Real-world systems that have characteristic input-output patterns but don't provide access to their internal states are as numerous as they are difficult to model. This dissertation introduces a modeling language for estimating and emulating the behavior of such systems given time series data. As a benchmark test, a digital violin is designed from observing the performance of an instrument. Cluster-weighted modeling (CWM), a mixture density estimator around local models, is presented as a framework for function approximation and for the prediction and characterization of nonlinear time series. The general model architecture and estimation algorithm are presented and extended to system characterization tools such as estimator uncertainty, predictor uncertainty and the correlation dimension of the data set. Furthermore a real-time implementation, a Hidden-Markov architecture, and function approximation under constraints are derived within the framework. CWM is then applied in the context of different problems and data sets, leading to architectures such as cluster-weighted classification, cluster-weighted estimation, and cluster-weighted sampling. Each application relies on a specific data representation, specific pre and post-processing algorithms, and a specific hybrid of CWM. The third part of this thesis introduces data-driven modeling of acoustic instruments, a novel technique for audio synthesis. CWM is applied along with new sensor technology and various audio representations to estimate models of violin-family instruments. The approach is demonstrated by synthesizing highly accurate violin sounds given off-line input data as well as cello sounds given real-time input data from a cello player.