Simulation of industrial knowledge mining algorithms using recurrent inference networks
David Al-Dabass, David J. Evans, K. Sivayoganathan · International Journal of Simulation and Process Modelling · 2006
Nodes in recurrent inference networks exhibit oscillatory characteristics even when the inputs are constant. This paper deals with estimating the causal parameters of such nodes. Hybrid recurrent nets combine arithmetic nodes and special integrator nodes to represent intelligent systems with oscillatory behaviour. Layers of hybrid nodes are arranged in hierarchies to model the complex output of such systems. Causal parameters are estimated from behaviour trajectories in a two stage process: time derivatives are determined first, followed by parameters. First order hybrid recurrent nets are employed to compute time derivates continuously as the behaviour is monitored. Further layers of arithmetic and hybrid nets then estimate the causal parameters of the complete model. Example applications are given to illustrate the techniques.