Deciding an appropriate level of abstraction, which would enable a model to answer some questions accurately while maintaining a reasonable level of complexity, is considered the most critical challenge in the modeling field in general. This thesis aims at presenting a new general approach for tackling the level of abstraction problem. The approach is investigated methodologically and practically. In the methodological part, a formalism of dynamic systems defining model components that form targets of abstraction is presented. A classification of the main categories of abstraction is also developed. A new taxonomy of the model abstraction techniques is then provided. Finally, a new automatic abstraction algorithm is developed. In the practical part, the ideas and algorithms presented are applied to Rule-Based MASs. A formal language for this type of models is defined and a set of five abstraction operators is developed. A framework is built implementing the approach and tested against known models. It has shown considerable success in replacing a detailed model with a set of simpler models that are suitable for each specific context.