Summarize the Execution Trace

At the end of a given problem-solving episode, there is potentially an overwhelming amount of information to process and much of it might not be relevant to future problem-solving episodes. Summarization processes allow the system to focus on the most significant aspects of a given cooperative activity, those parts that would take significant effort to reconstruct. Removing lesser details increases the generality of the resulting trace and can lessen later adaptive effort.

There are three principles that are used in summarizing the trace:

Keep

conversations that establish cooperation.

Keep

actions that coordinate cooperating agents.

Prune

actions that a dependency structure analysis reveals are reconstructible.

Communication must be included in the summary for the agents to establish that they are cooperating in future problem solving episodes. However, some communication is reconstructible, i.e. before doing a joint operator such as UNLOAD-TOGETHER, and so will be removed by the last principle.

Remembering points of coordination simplifies the cooperative interactions between agents in future problem solving episodes. Failure to remember points of coordination can lead to future difficulties even if both agents are working from the same problem-solving fragment. For example, if a heavy lifter reused a fragment composed of (LOAD HL BOX1 HANDTR1 ROOM1) and (LOAD HL BOX1 TRUCK1 STREET1), the planner will rederive the missing UNLOAD operator. However, without a WAIT operator, she might unload HANDTR1 before the hand-truck operator has a chance to push it to the street! To achieve coordination in the summarized trace, we use the following rule:

Convert agreed-to requests

into the corresponding WAIT operator.

Applying this rule to the above example will insert a WAIT operator between the two LOADs and ensure that the actions of the heavy lifter and the hand-truck operator are sensibly interleaved.

Reconstructability by the planner depends on the dependency structure implicit in the preconditions and effects of a given operator. For example, if a heavy lifter wishes to LOAD a box which is at ROOM1 onto TRUCK1 which is at STREET1, then the preceding LIFT and the CARRY operators can be reconstructed by backward chaining.

Figure 3 shows a summary of the execution trace from Figure 2 for each agent.