Learning the Behavior of Dynamical Systems form Examples

ABSTRACT This paper describes a general method for learning the behavior of dynamical systems. It can be used to control or predict their behavior. During a first step, the values of the state variables, and their interrelations are learned from the states which actually occur in the system. This data is organized in a topological map [Kohonen 84, 88]. Secondly, by looking at examples of the evolution of the system in time, the dynamical behavior is learned through a simple generalization process. This process constructs a vector field which describes the state transitions on the state space.