Algorithm Evolution with Internal Reinforcement for Signal Understanding.
Abstract
Automated program evolution has existed in some form for almost forty years. Signal understanding (e.g., signal classification) has been a scientific concern for longer than that. Generating a general machine learning signal understanding system has more recently attracted considerable research interest. First, this thesis defines and creates a general machine learning approach for signal understanding independent of the signal's type and size. This is accomplished through an evolutionary strategy of signal understanding programs that is an extension of genetic programming. Second, this thesis introduces a suite of sub-mechanisms that increase the power of genetic programming and contribute to the understanding of the learning technique developed. The central algorithmic innovation of this thesis is the process by which a novel principled credit-blame assignment is introduced and incorporated into the evolution of algorithms, thus improving the evolutionary process. This principled credit-blame assignment is done through a new program representation called neural programming and applied through a set of principled processes collectively called internal reinforcement in neural programming. This thesis concentrates on these algorithmic innovations in real world signal domains where the signals are typically large and/or poorly understood. This evolutionary learning of algorithms takes place in PADO, a system developed in this thesis for "parallel algorithm discovery and orchestration" and as a demonstrably effective strategy for divide-and-conquer in signal classification domains. This thesis includes an extensive empirical evaluation of the techniques developed in a rich variety of real-world signals. The results obtained demonstrate, among other things, the effectiveness of principled credit-blame assignment in algorithm evolution. This work is unique in three aspects.
Document Details
- Document Type
- Technical Report
- Publication Date
- Dec 05, 1998
- Accession Number
- ADA361241
Entities
People
- Astri Teller
Organizations
- Carnegie Mellon University