1 A Goal Oriented Strategy Based on Completion.pdfVIP

1A Goal Oriented Strategy Based on Completion Rolf Socher-Ambrosius* Max-Planck-Institut für Informatik, Im Stadtwald, D-W-6600 Saarbrücken, Germany email: socher@mpi-sb.mpg.de Abstract In this paper, a paramodulation calculus for equational reasoning is presented that combines the advantages of both Knuth-Bendix completion and goal directed strategies like the set of support strategy. Its soundness and completeness is proved, and finally the practical aspects of this method are discussed. 1 Introduction Knuth-Bendix completion has turned out to be a well suited tool for proving equational theorems. It seems, however, that this approach to equational reasoning does not support backward reasoning or goal directed strategies. Knuth-Bendix completion is a forward chaining method, deriving new conse- quences from the axioms and previously derived equations, intended to produce a complete set of equations and rewrite rules for the given equational specification. The proof of the theorem actually is more like a “side effect” of completion. The well-known problem with the set of support strategy [17] in combination with a paramodulation calculus [15] is the need to paramodulate into and from variables (see, for instance, [16]), both of which is precluded in Knuth-Bendix completion. The following clause set {f(a,b)≈a, a≈b, f(x,x)≈/x} which is taken from [16], is unsatisfiable. However, if the third clause is chosen as set of support, then a refutation can be obtained only by paramodulating into a variable. Snyder and Lynch [16] have found a way to overcome this difficulty. Their relaxed paramodulation calculus employs the set of support strategy, and it is complete without paramodulation into variables. Still, the relaxed paramodulation calculus cannot solve the problem with paramodulating from variables, as the following example (see [18]) shows: {(x*y)*z≈x*(y*z), x*e≈x, x*i(x)≈e, i(i(a))≈/a} (1) If the inequality is chosen as the set of support, then no refutation can be