A control strategy for expert systems is presented which is based on Shafer's Belief theory and the combination rule of Dempster. In contrast to well known strategies it is not sequentially and hypotheses-driven, but parallel and self organizing, determined by the concept of information gain. The information gain, calculated as the maximal difference between the actual evidence distribution in the knowledge base and the potential evidence determines each consultation step. Hierarchically structured knowledge is an important representation form and experts even use several hierarchies in parallel for constituting their knowledge. Hence the control strategy is applied to a layered set of distinct hierarchies. Depending on the actual data one of these hierarchies is chosen by the control strategy for the next step in the reasoning process. Provided the actual data are well matched to the structure of one hierarchy, this hierarchy remains selected for a longer consultation time. If no good match can be achieved, a switch from the actual hierarchy to a competing one will result, very similar to the phenomenon of restructuring in problem solving tasks. Up to now the control strategy is restricted to multi hierarchical knowledge bases with disjunct hierarchies. It is implemented in the expert system IBIG (inference by information gain), being presently applied to acquired speech disorders (aphasia).