Improving search through diversity

Abstract

Adding diversity to symbolic search techniques has not been explored in artificial intelligence. Adding a diversity criterion provides us with a powerful new mechanism for finding global maxima in complex search spaces and helps to alleviate the problem of premature convergence to local maxima. A theoretical analysis is presented of issues in diversity searching which previously haven't been addressed, and a domain-independent diversity-search algorithm for practical breadth-first searching is developed. Empirical results of an implementation in the CRESUS expert system for intelligent cash-management confirm that diversity can significantly improve the solution quality of symbolic searchers.