Evolutionary neural trees for modeling and predicting complex systems

Abstract Modeling and predicting the behavior of many technical systems is complicated because they are generally characterized by a large number of variables, parameters and interactions, and limited amounts of collected data. This paper investigates an evolutionary method for learning models of such systems. The models thus evolved are based on trees of heterogeneous neural units. The set of different neuron types is defined by the application domain, and the specific type of each unit is determined during the evolutionary learning process. The structure, size, and weights of the neural trees are also adapted by evolution. Since the genetic search used for training does not require error derivatives, a wide range of neural models can be constructed. This generality is contrasted with various existing methods for complex system modeling, which investigate only restricted topological subsets rather than the complete class of architectures. An improvement in the predictive accuracy and parsimony of models is reported, against backpropagation networks and other well-engineered polynomial-based methods for two problems: MacKey-Glass and Lorenz-like chaotic systems. The authors also demonstrate the importance of the selection pressure towards model parsimony for the improvement of prediction accuracy.