Gene regulatory networks (GRN) play a pivotal role in controlling cellular behavior. Learning these networks from experimental data is of great importance in biology. It requires proper computational methods and mathematical models to account for the complexity of the system dynamics. The combination of simplicity and ability to capture complex dynamics has made Boolean networks a popular modeling framework for gene regulatory networks. In this work, we present a novel particle swarm optimization (PSO) based technique for learning Boolean models of GRNs. To cope with the large solution space we use an improved variant of PSO, namely Dynamic Local Neighborhood based PSO (D-LPSO) and new techniques are introduced to handle binary data. We refer to the new algorithm as Binary D-LPSO, or BD-LPSO for short. We assess the performance of BD-LPSO on six regulatory systems of varying sizes and dynamics complexities. We find that this novel learning method is able to infer Boolean networks that captures the dynamics well and provide good predictive power. BD-LPSO also outperforms other state-of-the-art algorithms for learning Boolean networks, namely REVEAL and BestFit, in terms of accuracy, predictive power, and computational time.