This paper^1 discusses a novel approach for the construction of a context-free grammar based on a sequential processing of sentences. The construction of the grammar is based on a search algorithm for the minimum weight subgraph in an AND/OR graph. Aspects of optimality and robustness are discussed. The algorithm plays an essential role in a model for adaptive learning of probabilistic ordering. The novelty in the proposed model is the combination of well-established methods from two different disciplines: graph theory and statistics. The set-up of this paper is mainly theoretical, and we follow a quite formal approach. There is a close link with Optimality Theory, one of the mainstream approaches in phonology, and with graph theory. The resulting techniques, however, can be applied in a more general domain.