KALMAN FILTER NEURO-COMPUTING FOR BIOLOGICAL SYSTEM MODELS USING NEURAL NETWORKS

Abstract

The hierarchical neural network can be used to model biological systems such as plant growth, photosynthesis, evapotranspiration, etc. The development of back-propagation algorithm for neuron training has made it possible to use the layered network for simulating such non-linear systems. Modeling of such biological systems using the neural network often requires large number of layers and units in the network architecture because of the complexity of the system. The back propagation algorithm, however, often fails to achieve satisfactory identification of the system in the sense that output error minimization characteristics of the steepest descent scheme of the back propagation algorithm does not fit the problems involving large number of estimation parameters (synapse weights). An attempt of implementing Kalman filter algorithm (Kalman, 1960) in the procedure for training the neural network was made and evaluated. The performance of Kalman filter neuro-computing algorithm was compared to commonly used back propagation algorithm. Simulation of growth of radish sprouts under influence of changes in temperature and concentration of nutrient solution was attempted by two different neural network models, i.e., Kalman filter model and back propagation model. Results revealed a superior effectiveness of Kalman filter algorithm over the standard back propagation algorithm in the parameter estimation.