The final step of preparation is to decompose the improved, summarized trace into useful pieces. In the current design of the system, a useful piece is a plan fragment that corresponds to a portion of the trace which accomplishes some set of top-level goals. The steps of the trace are sequenced. Each step is then annotated by the top-level goals that it achieves.
Plan fragments are defined in two different ways. First, all actions that are annotated with a given goal G are part of the fragment created for goal G. Second, all actions in the trace that occur in the interval between the first time when an action annotated with G appears to when G is achieved are part of the fragment created for all goals that are completed during that time period.
Thus a given action can be in more than one plan fragment. The reason for the second technique is that it captures the interleaving of both dependent and independent goals; this feature is especially important because it remembers cooperative behavior that is not directly producible by the independent hierarchical planners of the agents.
Figure 5 shows three plan fragments derived from the HL trace shown in Figure 4. Note that HL has learned two simple plans for moving one box as well as the original three boxes.