Fuzzy genetics-based machine learning (FGBML) has frequently been used for fuzzy classifier design. It is one of the promising evolutionary machine learning (EML) techniques from the viewpoint of data mining. This is because FGBML can generate accurate classifiers with linguistically interpretable fuzzy if-then rules. Of course, a classifier with tens of thousands of if-then rules is not linguistically understandable. Thus, the complexity minimization of fuzzy classifiers should be considered together with the accuracy maximization. In previous studies, we proposed hybrid FGBML and its multiobjective formulation (MoFGBML) to handle both the accuracy maximization and the complexity minimization simultaneously. MoFGBML can obtain a number of non-dominated classifiers with different tradeoffs between accuracy and complexity. In this paper, we focus on heuristic rule generation in MoFGBML to improve the search performance. In the original heuristic rule generation, each if-then rule is generated from a randomly-selected training pattern in a heuristic manner. This operation is performed at population initialization and during evolution. To generate more generalized rules according to the training data, we propose new heuristic rule generation where each rule is generated from multiple training patterns. Through computational experiments using some benchmark data sets, we discuss the effects of the proposed operation on the search performance of our MoFGBML.