Machine Self-Teaching Methods for Parameter Optimization

摘要

The problem of determining near-optimum parameter-control logic is addressed for cases where a sensor or communication system is highly flexible and the logic cannot be determined analytically. A system that supports human-like learning of optimum parameters is outlined. The major subsystems are (1) a simulation system (described for a radar example), (2) a performance monitoring system, (3) the learning system, and (4) the initial knowledge used by all subsystems. The initial knowledge is expressed modularly as specifications (e.g., radar constraints, performance measures, and target characteristics), relationships (among parameters, intermediate measures, and component performance measures), and formulas. The intent of the learning system is to relieve the human from the very tedious trial-and-error process of examining performance, selecting and applying curve-fitting methods, and selecting the next trial set of parameters. A learning system to design a simple radar meeting specific performance constraints is described in detail, for experimental purposes, in generic object-based code. Keywords: Learning systems, System optimization, Radar, Control doctrine, Artificial intelligence.