Applicability of neuro-fuzzy function approximation in material-flow forecasting

Abstract

Production lines are often planned with very demanding time constraints. The offer phase commonly lasts only few days. It comprises conceptual work, layout planning, calculation and writing of the offer. Today, elaborated tools like material-flow simulation, which analyze the layout's quality, are rarely used in the offer phase, because they require much modeling time. Others are not applied due to strong restrictions or high abstraction level. In this paper, a neuro-fuzzy based approximation approach to forecast material-flow behavior in the offer phase is proposed. Within this approach, a neuro-fuzzy system maps material-flow simulation input parameters on performance measures. It is trained with results of multiple simulations using a model of the production line. Four different function approximation systems – CANFIS, NefPROX, Fuzzy Graph Construction and Adaptive Logic Networks – are tested. Results are compared to simple analytic forecast tools that are common in industry. The approximation approach yields results comparable to the analytic approach. Furthermore, it is able to learn behavior of production lines, which does not follow the analytic approach's model restrictions.