International Journal of Soft Computing and Engineering (IJSCE) ISSN: 2231-2307, Volume-2, Issue-4, September 2012 87 Abstract — This work provides the construction of Genetic Algorithm based Neural Network for parameter estimation of Fast Breeder Test Reactor (FBTR) Subsystem. The parameter estimated here is temperature of Intermediate Heat Exchanger of Fast Breeder Test Reactor. Genetic Algorithm based Neural Network is a global search algorithm having less probability of being trapped in local minimum problem as compared to Standard Back Propagation algorithm which is a local search algorithm. The various development stages of Genetic Algorithm based Neural Network such as the preparation of the training set, weight extraction from the genetic population, training of the neural network and validation phase etc have been described in detail. Keywords—Genetic Algorithm based Neural Network, Fast Breeder Test Reactor, Intermediate Heat Exchanger, Multi layer Perceptron I. INTRODUCTION For efficient and quick learning, the weight optimization of Back Propagation neural network has been carried out using Genetic Algorithm. Genetic Algorithms (GA) are adaptive search and optimization techniques, mimicking the principles of natural evolution. Genetic Algorithms have been proposed as one of the potential candidates for optimization of weight parameters of neural network. Conventionally, Standard Back Propagation network performing gradient descent learning algorithms have encountered difficulties of getting struck in local minima problem. Whereas Genetic Algorithm does not guarantee a global minimum solution, however it can locate the neighborhood of optimum solution much quicker than conventional strategies and provide encouraging results. This lessens the large number iterations needed for training the standard back propagation network too. Genetic Algorithms encodes the parameters of neural network as string of properties of the network, i.e. chromosomes. A large population of chromosomes representing many possible parameters sets is generated and crossover, mutation and reproduction are then performed in order to arrive at the best fit optimized parameters. Manuscript received on September, 2012 Subhra Rani Patra, Computer Division, EIRSG, Indira Gandhi Center for Atomic Research, Kalpakkam-603102, Tamil Nadu, India. R. Jehadeesan, Scientific Officer/F, at IGCAR, DAE, Kalpakkam, India. S.Rajeswari is Head of CSS, Computer Division, EIG,IGCAR, DAE, Kalpakkam, India. Genetic Algorithms work with population of individual strings, each representing a possible solution to the problem considered. Each string is assigned a fitness value accessing how good the solution is, to that particular problem. The string having high fitness values, participate in reproduction yielding new strings by cross breeding. The least fit individuals are discarded out. A whole new set of population, containing characteristics which are better than their ancestors, are generated by selecting the high fit individuals. Progressing in this way, after many generations, the entire population inheriting the best characteristics is formed. If the Genetic Algorithm is well implemented, the most promising areas of search space are explored, with the population having fitness values increasing towards the global optimum. A population is said to have converged if 95% of the individuals constituting the population share the same fitness value [1, 2]. The Fig.1 represents the flow chart representation of the Genetic Algorithm based Feed Forward Neural Network. Fig.1 Hybrid Genetic Algorithm for weight optimization of Artificial Neural Network The objective of this modeling is to evaluate primary and secondary sodium outlet temperatures for given mass flow rate in shell & tube side and respective inlet (primary and secondary) temperatures. The temperature prediction for severe unbalanced primary and secondary flow is performed using Nodal Heat Balance (NHB) method. Later it is modified with the help of Quadratic Upstream Interpolation for Convective Kinetics (QUICK) scheme and from QUICK code the required input data is generated for Artificial Neural Network modeling. The multilayer feed forward network model is observed to be best suited for parameter estimation in Intermediate Heat Exchanger. A comparison study of two different algorithms, back propagation and Genetic Algorithm based Neural Network is carried out which showed that Genetic Algorithm based Neural Network showed faster convergence with less number of iterations. Development of Genetic Algorithm based Neural Network model for parameter estimation of Fast Breeder Reactor Subsystem Subhra Rani Patra, R. Jehadeesan, S. Rajeswari, S. A.V. Satya Murty, M. Sai Baba