Adopting Relational Reinforcement Learning in Covering Algorithms for Numeric and Noisy Environments

Abstract

Covering algorithms (CAs) constitute a type of inductive learning for the discovery of simple rules to predict future activities. Although this approach produces powerful models for datasets with discrete features, its applicability to problems involving noisy or numeric (continuous) features has been neglected. In real-life problems, numeric values are unavoidable, and noise is frequently produced as a result of human error or equipment limitations. Such noise affects the accuracy of prediction models and leads to poor decisions. Therefore, this paper studies the problem of CAs for data with numeric features and introduces a novel non-discretization algorithm called RULES-CONT. The proposed algorithm uses relational reinforcement learning (RRL) to resolve the current difficulties when addressing numeric and noisy data. The technical details of the algorithm are thoroughly explained to demonstrate that RULES-CONT contribute upon the RULES family by collecting its own knowledge and intelligently re-uses previous experience. The algorithm overcomes the infinite-space problem posed by numeric features and treats these features similarly to those with discrete values, while incrementally discovering the optimal rules for dynamic environments. It is the first RRL algorithm that intelligently induces rules to address continuous and noisy data without the need for discretization or pruning. To support our claims, RULES-CONT is compared with 7 well-known algorithms applied to 27 datasets with four levels of noise using 10-fold crossvalidation, and the results are analyzed using box plots and the Friedman test. The results show that the use of RRL results in significantly improved noise resistance compared with all other algorithms and reduces the computation time of the algorithm compared with the preceding version, which does not use relational representation.