In the interest of brevity, the following is a summary of how the casebase of previous problem-solving knowledge can be maintained. The details of how CM uses prior problem-solving experience to learn cooperative procedures are given in [Garland & Alterman1995].
Updating a procedural knowledge casebase is too ambitious an undertaking for the agents to accomplish on-line. Off-line, agents have the time to analyze their performance in order to update their procedural knowledge casebase. Therefore, after the community of agents solves a problem, each agent evaluates her execution trace and stores it in CM after a three step process:
The purpose of the first step is to reduce the amount of information that is stored in memory (and thus simplify the remaining steps and later adaptive effort). The second step attempts to remove inefficiencies in agent behavior. The third breaks the execution trace into useful pieces of information. Those pieces of information are then placed in collective memory in a manner that makes them retrievable under similar circumstances in future problem-solving situations.
Cased based reasoning [Kolodner1993] can be a powerful mechanism for reducing communication in a multiagent domain, as has been previously demonstrated in [Sugawara1995, Ohko, Hiraki, & Anzai1995]. We expect that the reductions in communication shown in those works will carry over to our domain even though their benchmark systems used more communication. Furthermore, we expect CM to learn more efficient behavior for other operators as well.
CBR does have potential drawbacks, especially in a distributed setting, but they are manageable. First, there is a potential scaling problem, i.e. as the casebase gets larger, performance degrades because retrieval time increases. However, there are techniques to constrain the size of a casebase such as [Minton1990, Smyth & Keane1995]. Second, in a multiagent setting, global difficulties can arise when individuals use local criteria to determine the best case to retrieve as discussed in [NagendraPrasad, Lesser, & Lander1995]. In MOVERS-WORLD, any such inconsistencies are resolved in the exact same manner as when the plans were generated from first principles, i.e. discrepancies between these competing versions of cooperative behavior are resolved via communication.