An evolutionary computation approach for developing and optimizing discrete-time forecasting and classification models

CAFECS (Classification and Forecasting Evolutionary Computation System), a hybrid of genetic algorithms and Evolution Strategies, is introduced. Its purpose is to develop and optimize discrete-time forecasting and classification models, particularly in highly non-linear domains. Models are defined as sets of functions and are evaluated by a user-specified model criterion function that defines the extent to which one model is "better" than another. Model development is accomplished by searching a well-defined model domain by "evolving" a set of model domain subsets called colonies.;Principles derived from evolutionary theory and the system's distributed architecture are used to increase the system's performance. The role of natural selection in evolution is examined with an emphasis on the importance of cooperation. CAFECS permits a limited level of communication between colonies including the crossbreeding of models. The effects on system performance of different inter-colony communication strategies are examined. The gains obtained may be explained in part by Sewall Wright's shifting balance theory and the principle of sociogenesis.;Unlike other machine learning approaches, CAFECS is a data mining approach designed to work in conjunction with the user and apply user knowledge instead of exclusively deriving the models from relationships induced from data. However, this user knowledge is not required. The user can restrict the model domain or focus the search in areas which appear more promising and may improve system performance. Also, at any point in the development process, the user can enter new models that the user desires to test or has reason to believe may be useful. These models are then used by CAFECS in an attempt to create better models.;Among the tests performed, it was shown how CAFECS can be used to create hybrid models and how system performance was improved by evolving separate model colonies and permitting a limited exchange of models between the colonies. Also, it was shown that CAFECS can be used to discover attributes that enabled a decision tree induction algorithm to create a decision tree for classifying the terrain shown in a satellite image that is more accurate and smaller than normally obtainable.