In this study, we are concerned with a modularization of fuzzy relational equations that is converting a highly dimensional (multivariable) relational equation into a series of single input fuzzy relational equations. The problem originates from a need of handling (estimating) highly dimensional relational structures and is inherently associated with the curse of dimensionality present in relational fuzzy models. We propose a twolayer architecture and discuss a detailed optimization scheme leading to the determination of the fuzzy relations occurring there. Illustrative numerical studies are also included.1 Introduction: coping with multivariable fuzzy relation equations Fuzzy relational equations have been around for several decades almost from the inception of fuzzy sets [3, 4, 5, 11±18]. They generalize calculus of relations [8]. Without any exaggeration one may say that they can be used as a formal framework for numerous pursuits carried out in the setting of fuzzy sets, starting from various schemes of approximate reasoning (including a compositional rule of inference), models of decision-making, and relational structures in data mining [1,2,6,7,9,10,19]. A generic version of the equation arises in the formIn the above expression, A; B; C; . . . ; X are input fuzzy sets, R becomes a multidimensional fuzzy relation and Y denotes an output fuzzy set. The above topology of the equation can be viewed as a generic relational model of a multivariable static system. The convolution of the input fuzzy sets with the fuzzy relation can be realized in many possible ways; the s±t composition is a general option to work with while the well-known max±min composition comes as a special case. Generally, from the practical standpoint we are concerned with a collection of fuzzy relational equations (relational constraints) rather than a single equation. Then (1) generalizes to the form Yk Ak Bk Ck Á Á Á Xk R 2 with k 1; 2; . . . ; N.While the relational calculus supported by (1) or (2) exhibits a number of appealing features and is quite general, there are still open practical questions. They concern a way of solving such fuzzy relational equation(s). The approaches towards handling these equations fall under two general categories. First, analytical solutions hold under assumptions (e.g., a type of the convolution operation; in general we cannot come up with a general way of solving equations for the s±t composition). Second, if we con®ne ourselves to optimization methods (usually in the form of some iterative algorithms), it is very likely that these could maintain their effectiveness for relatively small dimensionality of the problem. What really hampers relational equations is the dimensionality of the problem that is cast in such setting. The fuzzy relation required to handle (1) grows up in size very quickly. This immediately raises an issue of ef®cient storage of the fuzzy relation itself. When proceeding with numeric optimization (estimation of R through some iterative computing), large fuzzy relations jeopardize the...