During the period of engagement, while
the agents are moving things about, the agents are
not concerned about finding an optimal solution -- any one will do. Later,
after the problem-solving episode,
the agents have time to reflect on their
activity and may be able to improve on their portion of the activity.
The input to this step is the summary of the activity trace. The basic idea is to go through the summary of the trace and attempt to find ways to improve the performance of the individual agent upon reuse of the summary. For example, in the case of the three boxes in the room, we would like the heavy lifter to independently discover that a better strategy is to load two boxes onto the hand-truck, rather than making an extra trip.
Two rules for improving summaries are:
Operators may be redundant before or after the application of the former rule. For example, the second and third PUSH-HANDTR and STAND-HANDTR operators are rendered redundant after re-sequencing. On the other hand, a LOAD-TOGETHER operator is redundant when it follows a WAIT for LOAD-TOGETHER operator regardless of re-sequencing.
Figure 4 shows the improved trace, using these rules, of the summarized execution trace shown in Figure 3. Note that the plan fragment is not yet optimal since the agents expect to unload the boxes in the same order that they loaded them instead of in the reverse order. However, it is more efficient than their original solution and will require extremely little adaptation to use in similar future problem-solving episodes.
There are dangers in attempting to improve an execution trace. Namely, any changes that are introduced may or may not turn out to be improvements. In the worst case, the changes introduce failure to a pattern of activity that was previously successful. Therefore we adopted a conservative policy towards introducing ``improvements." Nonetheless, our two rules were able to produce plan fragments that were more efficient than the original and were not directly producible by the planner.