Section 4. Adding Cooperative Procedures to Collective Memory

After the community of agents solves a problem, cooperative procedures are added to collective memory in order to improve the community performance in future problem-solving episodes. ADD-C-PROCS is the main procedure which an agent uses to update her procedural knowledge case-base with the information contained in her execution trace. This occurs in step 4 of Figure 1; pseudo-code is given in Figure 3. Briefly, the agent must summarize her run-time activities, both to remove unwanted data and to simplify further analysis. This summary is reorganized into chunks of operators that are associated with the goals they achieve. Then the operators in each chunk are modified to better prepare them for future re-use. Finally, the agent compares this chunk to current case-base entries to determine if the chunk should be added to the case-base.

  

Function ADD-C-PROCS(execution-trace)

  summary = SUMMARIZE-TRACE(execution-trace)
  foreach (chunk goals) in SPLIT-TRACE(summary)
    prepared = PREPARE-CHUNK(chunk)
    MODIFY-PKCB(prepared goals)

Figure 3: Algorithm to update collective memory.

In our model, the cooperative procedures stored in collective memory are culled primarily from execution traces. Others, such as [Carbonell1983] and [Laird, Rosenbloom, & Newell1986], have argued that an agent should store planning histories in memory. However, reusing plan derivations will not, in general, produce a sequence of actions to better solve a similar problem in the future. On the other hand, execution traces encapsulate the history of both planned and unplanned agent interactions with the domain. Consequently, cooperative procedures can be learned through the machinery of collective memory that were not developed in a single planning session.

Subsections:

• Summarization
• From Summary To Memory
  • SPLIT-TRACE.
  • PREPARE-CHUNK.
  • MODIFY-PKCB.