Throughout the practice of their joint activities, participants are learning and refining expectations about productive sequences of joint actions. Individual actors plan for new activities by borrowing from prior ones. The remembered success and failure of actions and decisions are the basis for improvement in behavior and the development of convention. Where initial experiences produce a set of biases, later ones refine them. New members of the community, or changes in the task environment, produce changes in the mix of regular joint behaviors that are continuously developing.
Our formulation of this issue ties learning to the functioning of the memory of the individual actors. After the community of actors solves a problem, individual actors retain in memory a description of that episode of joint behavior. This becomes the basis for the development of conventional behaviors.
Three critical features of the model we present of the individual's memory of prior coordinated behavior are:
There are two sets of reasons for why execution traces are the coinage of memory. The first has to do with the functioning of memory retrieval. Memory retrieval is tied to surface similarity features and not higher order relations between objects (see, for example, [24,26,27]). By tieing memory to execution traces, surface similarity features are likely to be present as indices to match against cues available in the task environment. The second set of reasons reflect emergent and situative aspects of joint behavior. An execution trace for a single actor encapsulates the history of both planned and unplanned activities within the domain. Consequently, an execution trace contains more information than any of the plans of the individual actor: the trace includes the history of actor behaviors required to solve the problem. This kind of information cannot be extracted from the planning process, but it is relevant in orienting the actor towards future joint behaviors.
Joint activities are a complex of coordination problems. Remembering points of coordination predisposes the individual actor to begin to anticipate those points of coordination in future related joint activities. Individuals coordinating their behavior must communicate at each coordination point in order to begin the next phase of joint activity. The points at which communication occurred between participants in the joint activity are stored as a part of the memory of prior activity. Maintaining this as a part of the individual actors expectations about the progress of a joint activity helps to maintain common ground while keeping the joint behavior coordinated.
If two actors are attempting to coordinate their behavior by remembering prior activity, their remindings should organize their current behavior so as to reduce work as the activity unfolds. Nevertheless, actors will assess the same situation in different manners. This reflects both differences in experience between actors and the open-endedness of interpretation in general. Since remembering is dependent on the actor's assessment of the situation, actors can retrieve incompatible plans. Communication provides an opportunity to redirect retrieval to harmonize the expectations of participating actors.
There are five parts to our discussion of how memory is updated after a session of activity:
Traces of run-time activities are cleaned to remove `noise', which simplifies further analysis and prevents reifying past mistakes. The cleaned trace is next reorganized into groups of actions that are related by the goals they achieve. Then each grouping is summarized and the remaining constituent actions are modified to derive a coordinated procedure that is prepared for future re-use. Optimizations may be performed. Finally, the actor compares the derived procedure to current casebase entries to determine if it should be added to the casebase.
CBR [45] has been applied to other systems with multiple actors. In particular, [36] used past run-time observations in a communication-free domain to supplement control strategies in order to reduce conflicts. Also, [59] learn to better allocate tasks among a community communicating via a contract net. In general, global difficulties can arise when individuals use local criteria to determine the best case to retrieve, as discussed in [57].