Teacher - an Automated System for Learning Knowledge-lean Heuristics

In this report, we present TEACHER, a system for learning knowledge-lean heuristics under resource constraints. TEACHER is an implementation of a genetics-based learning framework we have dev eloped for improving the performance of heuristics in application problem solvers. Besides providing a flexible and modular framework for conducting experiments, TEACHER provides (a) a test-bed for experimenting with various resource scheduling, generalization, and heuristics-generation strategies, (b) an automated learning system that can be easily interfaced to new applications and can be customized based on user requirements and target environments. This report describes the application-independent (AppI) functions provided by TEACHER, and the applicationdependent (AppD) functions for interfacing to new problem solvers. By adjusting various global parameters in TEACHER, users can control the numerous options and alternatives in TEACHER. To illustrate our design, we use CRIS, a genetics-based test-pattern generation system, as a running example throughout the report.