Performance analysis of the LAMDA fuzzy algorithm improvements in different case studies

Abstract

Learning Algorithm for Multivariable Data Analysis (LAMDA) is a fuzzy approach, which has been used in clustering and classification processes. Recently, extensions have been proposed for LAMDA, to improve its performance in classification tasks. The first one is called LAMDA-FAR, which proposes a new criterion to validate functional states after recognition, based on the minimum and maximum calculated distances between the two membership degrees with the highest values. The second extension is called LAMDA-HAD, which proposes two strategies to improve LAMDA performance. The first strategy calculates an adaptive Global Adequacy Degree (GAD) of the Non-Informative Class (NIC) to each class to prevent that correctly classified individuals will be assigned to the NIC class. The second strategy calculates the similarity among the GAD of an individual and all ones of each class, to make a more reliable assignment. This article analyzes the performance of these techniques for different classification problems. The goal is to define the application context for each one. Each case study was defined by a set of data in an operational context, which must be used by the classification techniques to obtain accurate results. LAMDA-HAD was better with unbalanced classes, while LAMDA-FAR was excellent for discovering new classes. Both algorithms worked well for different levels of noise (which can represent faults in the sensors), a factor important in diagnostic tasks. The aim of this paper is to determine the correct utilization profile of each LAMDA technique adjusted to the properties of the problems under study.