Flat-structured Problems (FP) are an important class of Cooperative Distributed Problem Solving applications which include speech understanding, vehicle monitoring, transport dispatching and so on. So far, a number of approaches to FPs have been developed such as those in Hearsay-II and DVMT. Most of these approaches use predictions or goals to guide bottom-up problem solving. However, most predictions and goals in these approaches are based on local view of problem solving states. Although the improved architecture of DVMT allowed a high-level view, no explicit algorithm was given. This paper gives an integrated approach to FPs which makes top-down predictions from global problem solving states, guides bottom-up solving by predictions, verifies and modifies predictions by newly-created hypotheses, and guide bottom--up solving once again. This approach not only enhances the directing role of predictions obtained from global problem solving states, but also makes problem solving flexible due to the prediction verification mechanism.