Hybrid Soft Computing for Classification and Prediction Applications

Abstract

Soft computing (SC) is an association of computing methodologies that includes as its principal members fuzzy logic (FL), neural computing (NC), evolutionary computing (EC), and probabilistic computing (PC). These methodologies allow us to deal with imprecise, uncertain data, and incomplete domain knowledge that are encountered in real-world applications. We will describe the advantages of using SC techniques, and in particular we will focus on the synergy derived from the use of hybrid SC systems. This hybridization allows us to integrate knowledge-based and data-driven methodologies to construct models for classification, prediction, and control applications. In this presentation we will describe three real-world SC applications: the prediction of time-to-break margins in paper machines; the automated underwriting of insurance applications; and the development and tuning of raw-mix proportioning controllers for cement plants.

The first application is based on a model that periodically predicts the amount of time left before an unscheduled break of the web in a paper machine. The second application is based on a discrete classifier, which assigns a vector of real-valued and attribute-values inputs, representing an insurance applicant’s vital data, to a rate class, representing the correct insurance premium. The third application is based on a hierarchical fuzzy controller, which determines the correct proportion of the raw material to maintain certain properties in a cement plant.

The similarity among these applications is the common process with which their models were constructed. In all three cases, we held knowledge engineering sessions (to capture the expert knowledge) and we collected, scrubbed and aggregate process data (to define the inputs for the models). Then we encoded the expert domain knowledge using fuzzy rule-based or case-based systems. Finally, we tuned the fuzzy system parameters using either local or global search methods (NC and EC, respectively) to determine the parameter values that minimize prediction, classification, and control errors.