A double-layered learning approach to acquiring rules for classification: Integrating genetic algorithms with similarity-based learning

In this paper, we describe a machine learning technique based on a double-layered architecture and Genetic Algorithms (GAs), which can be used to learn decision rules for financial classification. Once the rules have been acquired, they can be stored in an expert system for future application. Genetic algorithms represent a class of learning algorithms modeled on the biological evolution process. Equipped with unique search behavior and solution-seeking properties, they provide an interesting technique for such classification tasks as bankruptcy prediction and credit analysis. However, some modifications to the basic genetic algorithms are necessary in order to make the method suitable for solving these classification problems. One of the objectives of this paper is to develop a representation scheme for the concepts to be learned that can be incorporated in a genetic algorithm. More importantly, we expand on the concept of the genetic algorithm and develop a learning method called the Double-layered Learning System (DLS) that integrates the genetic algorithm with a similarity-based learning technique called the Probabilistic Learning System (PLS1). By combining a learning paradigm that searches through the ‘hypothesis space’ (GA), with a paradigm that searches through the ‘instance space’ (PLS1) DLS permits synergistic improvement of the learning performance. We demonstrate the feasibility of such a hybrid approach computationally by presenting the design, implementation, and performance evaluation of DLS. DLS proves to be an effective improvement over both GA and PLS1. The efficacy of the results points towards the importance of such hybrid approaches in providing more robust machine learning methods.