A Connectionist Approach for Learning Search - Control Heuristics forAutomated Deduction SystemsChristoph

b b b b b b b b b b b b b b b b b b b Abstract The central problem in automated deduction is the explosive growth of search spaces when proof length increases. In this paper, a con-nectionist approach for learning search-control heuristics for automated deduction systems is presented. In particular, we show how folding architecture networks, a new type of neural networks capable of solving supervised learning tasks on structured data, can be used for learning heuristics evaluation functions for algebraic (logical) expressions and how these evaluation functions can then be used to control the search process for new proof problems. Experimental results with the automated deduction system Setheo in an algebraic domain show a considerable performance improvement. Controlled by heuristics which had been learned from simple problems in this domain the system is able to solve several problems from the same domain which had been out of reach for the original system .