Applying Bacterial Memetic Algorithm for Training Feedforward and Fuzzy Flip-Flop based Neural Networks László Gál 1,2 János Botzheim 3,4 László T. Kóczy 1,4 Antonio E. Ruano 5 1 Institute of Information Technology and Electrical Engineering, Széchenyi István University, Gyr, Hungary 2 Department of Technology, Informatics and Economy, University of West Hungary Szombathely, Hungary 3 Department of Automation Széchenyi István University Gyr, Hungary 4 Department of Telecommunication and Media Informatics Budapest University of Technology and Economics, Budapest, Hungary 5 Centre for Intelligent Systems, FCT, University of Algarve, Portugal Email: gallaci@ttmk.nyme.hu, {botzheim, koczy}@{sze, tmit.bme}.hu, aruano@ualg.pt Abstract—In our previous work we proposed some extensions of the Levenberg-Marquardt algorithm; the Bacterial Memetic Algorithm and the Bacterial Memetic Algorithm with Modified Operator Execution Order for fuzzy rule base extraction from input- output data. Furthermore, we have investigated fuzzy flip-flop based feedforward neural networks. In this paper we introduce the adaptation of the Bacterial Memetic Algorithm with Modified Operator Execution Order for training feedforward and fuzzy flip- flop based neural networks. We found that training these types of neural networks with the adaptation of the method we had used to train fuzzy rule bases had advantages over the conventional earlier methods. Keywords— Bacterial Memetic Algorithm, Fuzzy Flip-Flop, Levenberg-Marquardt method, Neural Network. 1 Introduction Bacterial type evolutionary algorithms are inspired by the biological bacterial cell model [1,2]. The Bacterial Memetic Algorithm (BMA) is a recent method for fuzzy rule base extraction from input-output data for a certain system [7]. We have investigated its properties intensely and found some points where its performance in the fuzzy rule base identification could be improved. The recent bacterial type algorithms we proposed were named Bacterial Memetic Algorithm with Modified Operator Execution Order (BMAM), Improved Bacterial Memetic Algorithm (IBMA) and Modified Bacterial Memetic Algorithm (MBMA) [3,4]. They are both memetic algorithms and apply alternatively global and local search for identifying fuzzy rule bases from input-output data automatically when no human expert to define the rules is available. Neural Networks belong to the Soft Computing area like Fuzzy Systems and Evolutionary Computing. They can be used for modeling a certain system where input-output data pairs exist. The neural networks are inspired by biological phenomena: the brain itself and other parts of the neural system. Fuzzy Flip-Flops are extended forms of the binary flip-flops that are widely used in digital technics [5]. They use fuzzy logic operations instead of Boolean logic ones and require fuzzy inputs, furthermore they produce fuzzy outputs instead of digital values. Our previous works were developing the Bacterial Memetic Algorithm applied for fuzzy rule base identification (FRBI) and investigating various types of Fuzzy Flip-Flops (F 3 ) used in feedforward neural networks (FFNN) as replacements of the neurons [6]. We trained the Fuzzy Flip-Flop based Neural Networks (FNN) with the Levenberg-Marquardt (LM) based training method as it is a widely used and accepted one. However, we faced the same problems with the LM based feedforward neural network training as in the fuzzy rule base identification. Therefore we have adopted the Bacterial Memetic Algorithm with Modified Operator Execution Order (BMAM) for training Neural Networks. Our goal was to improve the learning capabilities of feedforward neural networks with a bacterial type evolutionary approach. In this paper we propose the adaptation of the BMAM for training feedforward neural networks, and we study and evaluate the respective results. From another aspect another paper was proposed here where we report on the findings of our investigations of the properties of different types of FNNs trained with BMAM [14]. 2 Bacterial Memetic Algorithm with Modified Operator Execution Order (BMAM) 2.1 Bacterial Memetic Algorithm (BMA) Bacterial Memetic Algorithm (BMA) is a very recent approach used for fuzzy rule base identification (FRBI) [7]. It combines global and local search. For the global search it uses bacterial type evolutionary approach and for the local search the Levenberg-Marquardt method is deployed. Previous work confirmed that the Pseudo-Bacterial Genetic Algorithm (PBGA) and the Bacterial Evolutionary Algorithm (BEA) were rather more successful in this area than the conventional genetic algorithms [1,2]. 2.1.1 Bacterial mutation PBGA is a special kind of Genetic Algorithm (GA) [8], it introduces a new “genetic” operation called bacterial mutation. For the algorithm, the first step is to determine how the problem can be encoded in a bacterium (chromosome). In case of modelling fuzzy systems the ISBN: 978-989-95079-6-8 IFSA-EUSFLAT 2009 1833