Logical operation based fuzzy MLP for classification and rule generation

A jilzzy layered neural network/or classification and rule generation is proposed using logical neurons. II can handle uncertainty and/or impreciseness in the input as well as the output. Logical operators, namely, t­ norm T and t-conorm S involving And and Or neurons, are employed in place of the weighted sum and sigmoid functions. Various/uzzy implication operators are introduced to incorporate different amounts ofmutual interaction during the back propagation Q{ errors. In case 0/partial inputs the model is capable of querying the user/or the more important/eature information, (f and when required. Justification for an inferred decision may be produced in rule/orm. The built-in And-Or structure of the network enables the generation Qf appropriate rules expressed as the disjunction Qf conjunctive clauses. The e.ffectiveness o.fthe model is tested on a speech recognition problem and on some art({tcially generated pattern sets. 1. INTRODUCTION teacher. The learning procedure has to determine the internal parameters of the hidden units based on its Artificial neural networks or connectionist models knowledge of the inputs and desired outputs. They are designed perhaps as an attempt to emulate lies in their capability in modeling uncertain or am­ biguous data often encountered in real life. There have human performance and function intelligently. An ad­ been several attempts recently (Bezdek & Pal, 1992; vantage of neural net lies in their high computation rate provided by massive parallelism, so that real-time in making a fusion of fuzzy logic and neural processing of huge data sets becomes feasible with networks for better performance in decision making proper hardware, Information is encoded among the various connection weights in a distributed manner. systems. The multilayer perceptron (MLP) (Rumelhart & The present work is an attempt in this regard and McClelland, 1986; Hinton, 1987) consists of multiple describes a logical version ofthe feedforward MLP using layers of simple, sigmoid processing elements (nodes) the concept offuzzy sets at various stages, The proposed or neurons that interact using weighted connections. model represents the input vector in terms oflinguistic