F. Progress in the Solving of a Circuit Design Problem.pdfVIP

Progress in the solving of a circuit design problem Laurent Granvilliers and Fr´ed´eric Benhamou IRIN, Universit´e de Nantes B.P. 92208, F-44322 Nantes Cedex 3, France {Laurent.Granvilliers,Frederic.Benhamou}@irin.univ-nantes.fr Abstract. A new branch and prune algorithm for globally solving nonlinear sys- tems is presented. It combines a multidimensional interval Newton method with the constraint satisfaction algorithm HC4 [1]. The main contributions of this paper are the fine-grained interaction between both algorithms that avoids some unnecessary computation, and the description of HC4 in terms of a chain rule for constraints’ projections. The new algorithm is experimentally compared with the global methods from Ratschek and Rokne [14] and from Puget and Van Hentenryck [13] on a circuit design problem from Ebers and Moll [5]. An interval enclosure of the solution with a precision of twelve significant digits is computed in four minutes, providing an improvement factor of five on the same machine. Keywords: Constraint satisfaction, interval analysis, Newton method, automatic differentiation, circuit design. 1. Introduction and related work Interval methods for solving nonlinear systems were adapted from nu- merical analysis techniques in the last thirty years. One may cite Tay- lor expansions, Newton methods and Gauss-Seidel iterations. In the last decade, some authors (Cleary [4], Hyv¨onen [9], Older [12], Ben- hamou [2], Lhomme [10], Faltings [6], Van Hentenryck [15]) have shown that constraint satisfaction techniques are a powerful alternative to traditional interval methods. One of the most efficient solver, Numeri- ca [16], is based on a combination of local consistency (domain pruning) and bisection-based search. The core algorithm of Numerica is a branch and prune algorithm where the