We present a machine learning methodology (models, al- gorithms, and experimental data) to discovering the agent dynamics that drive the evolution of the social groups in a community. We use a parameterized probabilistic agent- based model integrating with micro-laws to present the agent dynamics. The micro-laws with different parame- ters present different actors' behaviors. Our approach is to identify the appropriate parameters in the model including discrete parameters together with continues parameters. To solve this mixed optimization problem, we develop heuris- tic expectation-maximization style algorithms for determin- ing the appropriate micro-laws of a community based on either the observed social group evolution, or observed set of communications between actors without considering the semantics. Also, in order to avoid the resulting combina- torial explosion, we appropriately approximate and opti- mize the objective within a coordinate-wise gradient ascent (search) setting for continuous (discrete) variables. Finally, we present the learning performance from extensive experi- ments.