Classification is a key problem in the machine learning field, and some metaheuristics have been successfully adapted to answer this problem. However, difficulties commonly arise when a classi- fication problem is described by numerical attributes, which are very common in most real-world tasks. Therefore, a metaheuristic-based classification algorithm often needs to be adapted to support this new type of attributes. In this study, we propose a method for representing and evolving classifi- cation rules with numerical attributes and extend the MOCA-I classification algorithm to support this type of attributes. We investigate several variants for the neighborhood generation mechanism that is at the core of the local search process, and propose two improvements on the general algorithm. Experimentations are done to evaluate the performance of each of the proposed variants.