Hierarchical Learning of Complex Systems.
Abstract
A central problem in forecasting and controlling nonlinear processes is quantifying the trade-off between available computational resources, model complexity, and prediction error. A more subtle, but important issue that strongly affects success is the choice of representation, or model class. Should one use Fourier or wavelet transforms, neural networks, hidden Markov models, or fuzzy logic, as modeling frameworks? As a tool for answering the questions of representation dependence, brittleness, and resource requirements, we introduced hierarchical e-machine reconstruction. This led to a number of detailed analyses of intrinsic computational capability in low-dimensional and spatially-extended nonlinear dynamical systems. This Final Technical Report outlines our investigations of the computational mechanics of learning complex systems during the period beginning 1 April 1991 and ending 29 February 1996. This project was supported under AFOSR grant number 91-0293. The report reviews the activities, personnel, and research highlights and lists the published papers and those currently under review.
Document Details
- Document Type
- Technical Report
- Publication Date
- Feb 29, 1996
- Accession Number
- ADA312476
Entities
People
- Donald A. Glaser
Organizations
- University of California, Berkeley