An agent-based approach for predictions based on multi-dimensional complex data

This paper presents an agent-based approach to the identification of prediction models for continuous values from multi-dimensional data, both numerical and categorical. A simple description of the approach is: a number of agents are sent to the investigated data space; at the micro-level, each agent tries to build a local linear model with multi-linear regressions by competing with others; then at the macro-level all surviving agents build a global model by introducing membership functions. Three tests were carried out and the performance of the approach was compared with that of a neural network. The results of the three tests show that the agent-based approach can achieve good performance for some data sets. The approach complements rather than competes with other Soft Computing methods.