Controlling search with meta-Brave

Abstract

We outline the Brave system: a Horn-clause language Brave, a dialect of Prolog designed for OR-parallel execution, plus meta-Brave, containing extra-logical features. Brave has been stripped of Prolog's cut, assert and retract, enforcing a declarative programming style which enables easier parallelisation. However, we need replacements for those predicates, in particular to code parallel versions of search algorithms, where guidance of the search in one subtree requires information from another subtree. Meta-Brave features directives which allow algorithmic knowledge about the domain to be stated independently from the Brave application program. This guidance is provided in three main ways: partial results are selectively remembered (lemmas), conditions of early failure of subproofs are stated (pruning), the style of search is altered (strategies). We show how this combination of features makes the Brave system specially suited to writing clear programs, for example search algorithms taken from Artificial Intelligence. In this paper we investigate how OR-parallelism can be exploited in hill-climbing, genetic algorithms and best-first search. By simulation we analyse the parallelism obtained.