Learning in First-Order Logic using Greedy Evolutionary Algorithms

In evolutionary computation ‘learning’ is a byproduct of the evolutionary process as successful individuals are retained through stochastic trial and error. This learning process can be rather slow, due to the weak strategy used to guide evolution. A way to overcome this drawback is to incorporate greedy operators in the evolutionary process. This paper investigates the effectiveness of this approach for inductive concept learning in (a fragment of) First-Order Logic (FOL). This is done by means of a new greedy evolutionary algorithm. The algorithm evolves a population of Horn clauses. Randomized greedy operators are employed for generalizing and specializing a clause. The degree of greediness of each operator is determined by a parameter. In this way, the user can control the greediness of the learning process by setting the parameters to specific values. A typical case study in Inductive Logic Programming (the KRK endgame problem) is used for testing the learning method. The effect of the greedy operators on the learning process is analyzed, by means of extensive experiments with different values of their parameters. Moreover, the robustness of the method to noise in the training examples is investigated.