Knowledge mining architectures using recurrent hybrid inference networks

Hybrid recurrent nets combine arithmetic and integrator elements to form nodes for modelling the complex behaviour of intelligent systems with dynamics. Given the behaviour pattern of such nodes it is required to determine the values of their causal parameters. The architecture of this knowledge mining process consists of two stages: time derivatives of the trajectory are determined first, followed by the parameters. Hybrid recurrent nets of first order are employed to compute derivatives continuously as the behaviour is monitored. A further layer of arithmetic and hybrid nets is then used to track the values of the causal parameters of the knowledge mining model. Applications to signal processing are used to illustrate the techniques.