Association rule discovery is an ever increasing area of interest in data mining. Finding rules for attributes with numerical values is still a challenging point in the process of association rule discovery. Most of popular methods for association rule mining cannot be applied to the numerical data without data discretization. There have been efforts to resolve the problem of dealing with numeric data. These approaches suffer from problems which are discussed in this paper. This work proposes a multi-objective genetic algorithm approach for mining association rules for numerical data. Several measures are defined in order to determine more efficient rules. Three measures, confidence, interestingness, and comprehensibility have been used as different objectives for our multi objective optimization which is amplified with genetic algorithms approach. Finally, the best rules are obtained through Pareto optimality. This method is based on the notion of rough patterns that use rough values defined with upper and lower intervals to represent a range or set of values. Mutation and crossover operators give a powerful exploration ability to the method and allow it to find out the best intervals of existing numerical values. The experimental results show that the generated rules by this method are more appropriate - based on several different characteristics - than the similar approaches' results, and our method outperforms these methods.