Development of accurate and reliable models for predicting the strength of rocks and rock masses is one of the most common interests of geologists,civil and mining engineers and many others.Due to uncertainties in evaluation of effective parameters and also complicated nature of geological materials,it is difficult to estimate the strength precisely using theoretical approaches.On the other hand,intelligent approaches have attracted much attention as novel and effective tools of solving complicated problems in engineering practice over the past decades.In this paper,a new method is proposed for mining descriptive Mamdani fuzzy inference systems to predict the strength of intact rocks and anisotropic rock masses containing well-defined through-going joint.The proposed method initially employs a genetic algorithm(GA)to pick important rules from a preliminary rule base produced by grid partitioning and,subsequently,selected rules are given weights using the GA.Moreover,an information criterion is used during the first phase to optimize the models in terms of accuracy and complexity.The proposed hybrid method can be considered as a robust optimization task which produces promising results compared with previous approaches. Development of accurate and reliable models for predicting the strength of rocks and rock masses is one of the most common interests of geologists, civil and mining engineers and many others. Due to uncertainties in evaluation of effective parameters and also complicated nature of geological materials, it is difficult to estimate the strength precisely using theoretical approaches.On the other hand, intelligent approaches have attracted much attention as novel and effective tools of solving complicated problems in engineering practice over the past decades. In this paper, a new method is proposed for mining descriptive Mamdani fuzzy inference systems to predict the strength of intact rocks and anisotropic rock masses containing well-defined through-going joint. The proposed method initially employs a genetic algorithm (GA) to pick important rules from a preliminary rule base produced by grid partitioning and , subsequently, selected rules are given weights using the GA. Moreover, an information crit erion is used during the first phase to optimize the models in terms of accuracy and complexity.The proposed hybrid method can be considered as a robust optimization task