Relational Reinforcement Rule Induction and the Effect of Pruning

摘要

Covering Algorithm (CA) is a Machine Learning field that produces a powerful repository represented as simple if-then rules. Although this field is well established with discrete data but it has its deficiency when dealing with numeric data. This paper introduces a new algorithm called RULES-CONT, which deal with continuous attributes using Relational Reinforcement Learning (RRL). This algorithm is a non-discretization algorithm that directly deal with all kind of data and transfer past experience to improve and generalize its learning. Different pruning levels will also be tested with RULES-CONT, in order to study the possibility of using pruning with RRL to reduce the time and solve the speed problem. The main contribution is to propose a novel solution for rule induction and analyse the most effective pruning level that should be integrated with RRL. Several experiments will be presented to compare RULES-CONT with other algorithms. These experiments are validated using 10-cross-fold validation and Friedman test to measure the significance between the algorithms and decide on the most suitable pruning level with RRL.