Incremental Knowledge Acquisition for Improving Probabilistic Search Algorithms

Abstract

A new incremental knowledge acquisition approach for the effective development of efficient problem solvers for combinatorial problems based on probabilistic search algorithms is proposed. The approach addresses the known problem of adapting probabilistic search algorithms, such as genetic algorithms or simulated annealing, by the introduction of domain knowledge. This is done by incrementally building a knowledge base that controls parts of the probabilistic algorithm, e.g. the fitness function and the mutation operators in a genetic algorithm.

The probabilistic search algorithm is monitored by a human who makes recommendations on search strategy based on individual solution candidates. It is assumed that the human has a reasonable intuition of the search problem. The human adds rules to a knowledge base describing how candidate solutions can be improved, or characteristics of candidate solutions which he/she feels are likely or unlikely to lead to good solutions. Our framework is inspired by the idea of (Nested) Ripple Down Rules where humans provide exception rules to rules already existing in the knowledge base using concrete examples of inappropriate performance of the existing knowledge base.

We present experiments on industrially relevant domains of channel routing as well as switchbox routing in VLSI design. We show very encouraging inital experimental results demonstrating that our approach can solve problems comparably well to other approaches. These other approaches use algorithms developed over decades, while we were able to develop an effective search procedure in a very short time. A brief discussion outlines our KA experience with these experiments.