Execution Traces

In our model, the data prepared and stored in collective memory are segments of execution traces. Others (Veloso & Carbonell, 1993; Laird, Newell, & Rosenbloom, 1986) have argued for what Carbonell (1983) refers to as the derivational history model for storage in memory. The notion is that an agent stores in memory a history of the planning session that can be replayed in future planning episodes. Our system stores execution traces rather than derivational histories because in uncertain worlds a derivational history will not adequately reflect the constant tinkering that is done at runtime to accomplish a multi-agent task -- and an execution trace will. At runtime, an agent's activity may not go as expected for a variety of reasons, e.g.,

Communication.

She could be interrupted by requests from other agents.

Conflict.

She could have to contend with other agents over a limited resource.

Infidelity.

Another agent could fail to keep a commitment.

Because of this, the execution trace contains more information than any of the plans of the individual agents; the trace includes the history of agent interaction required to solve the problem. This kind of information cannot be extracted from the planning process.