BioHEL: Bioinformatics-oriented Hierarchical Evolutionary Learning

This technical report briefly describes our recent work in the iterative rule learning approach (IRL) of evolutionary learning/genetics-based machine learning. This approach was initiated by the SIA system [12] A more recent example is HIDER [1]. Our approach integrates some of the main characteristics of GAssist [4], a system belonging to the Pittsburgh approach of Evolutionary Learning, into the general framework of IRL. Our aims in developing this system are use all the good characteristics of GAssist but at the same time overcome some of the scalability limitations that it presents. The document is splitted in five parts, knowledge representation, general workflow, fitness function, some illustrative results and further work. The system has the following characteristics: • Each individual is a single rule, and each GA run learns one rule at a time. • In order to learn all the rules of a domain this process is applied iteratively. At the end of each iteration the examples covered by the rule that have been learned are discarded from the training set. In this way, the GA is forced to explore other areas of the search space.