Structuring the RM into a pipeline of five layers certainly implies that the flow of information in an IMMPS goes in line with the pipeline view. At a first glance, this is actually the case. Presentation goals and commands form the triggering input to an IMMPS. While being processed through the layers, goals/commands are eventually transformed into presentations composed of media objects to be displayed to the user. Thus the flow of information during this transformation is primarily ``top-down''. On the other hand, it must be stressed that many interactions can occur between the layers. As indicated by the notification arrows in Fig. 12 the RM supports bidirectional interactions between layers, too. Each layer can notify the result of its processing to the layer above, possibly together with some additional information, such as causes of failure, recovery strategies, explanations, etc.. For example, the task division between Goal Selection and Goal Refinement requires bidirectional information flow between the Control Layer and the Content Layer (cf. subsections 4.1 and 4.2). Another example is the generation of cross-media reference expressions. To make a textual reference to the location of a picture or a picture part in a presentation, the Content Layer needs to know the spatial attributes of the particular location. However, these attributes may become available only after the layout has been realized. Thus the Content Layer has to wait for a notification of the Layout Realization component before it can accomplish the attribute selection. Generally speaking, bidirectional communication between layers becomes necessary when decision making in a layer also depends on decisions to be made in one of the layers below.
