The objective of this research is to apply grammatical inference models recently developed in the field of linguistics and phonology, as a basis for abstraction, composition, symbolic control, and learning in distributed multi-agent cyber-physical systems. The approach is to map the system dynamics, specifications, and task interdependences to finite abstract models, and then describe the desired behavior of the system in an appropriate grammar that can be decomposed into local agent specifications. In this framework, the agents can learn the behavior of their environment by observing its dynamics, and update their specifications accordingly. The proposed approach to learning in cyber-physical systems, which is based on grammatical inference at a purely discrete level, is a significant departure from current works. Following this approach, one can reason about large-scale processes resulting from event interdependencies between agents, without having to construct large product systems. To realize this plan, specific technical advances on modeling, abstraction, and control synthesis are proposed. Questions related to formally factoring and composing heterogeneous systems are pervasive in the fields of formal languages and computational learning. There are also applications of commercial significance in the area of discovering new azeotropic mixtures based on documented pairs of compounds that are known to have the particular property. Proposed dissemination and outreach activities include the involvement of middle and high school students and teachers, integrated in existing NSF-sponsored programs at the University of Delaware and Boston University.