Policy-based network management (PBNM) is a promising but not yet delivering discipline aimed at automating network management decisions based on expert knowledge and strategic business objectives. One of the issues scarcely addressed in PBNM is the stability of the managed system as the result of the dynamic interaction between the ?natural? network behaviour and the autonomous management decision making. Yet this issue is central to the design of a self-management networking system comprised of autonomous entities making decisions driven by policies with often unknown consequences. Decisions made by one entity may change the context and configuration of other autonomous entities which may in turn react changing the context and configuration of the first entity triggering an unbounded chain of re-configuration actions. It is possible to model obligation policies and their constraints with finite state transducers (FST). It is also possible to learn patterns of recurrent behaviour using Bayesian networks (BN), a structurally similar kind of graph. The method presented in this paper analytically composes both finite state machines to derive predictions of the consequences of enforcing a given policy improving system stability.