Section 2. Overview of Collective Memory

Humans provide a natural model of pragmatic agents situated in a multi-agent world. [Cole & Engeström1993] argues that the development of distributed cooperative behavior in people is shaped by the accumulated cultural-historical knowledge of the community. Our learning techniques are motivated by this argument and use a structure called collective memory to store the breadth of knowledge the community acquires through interacting with each other and the world during the course of solving sequences of distinct problems. In this paper, we assume a distributed (rather than centralized) memory model wherein each agent maintains only a memory of her own interactions; jointly, these memories represent the collective memory of the community.

The cornerstone of collective memory is a cooperative procedures case-base that augments the agents' first-order planner. In other words, this work extends single agent second-order planners [Alterman1988,Hammond1990,Veloso & Carbonell1993] into multi-agent domains. The procedures stored in collective memory will be effective in future problem-solving episodes if they capture regularities in their domain of activity.

  

1. Give the community a set of goals to achieve.
2.  Allocate the goals among the agents.
3.  Until all agents have achieved their goals:
   • 3a. Active agents use collective memory to create or adapt a plan.
   • 3b. Agents with plans attempt an operator.
4.  Add cooperative procedures to collective memory.

Figure 1: One cycle of community activity.

We consider the task environment for collective memory to be problem-solving by a community of multiple adaptive agents with differing roles and abilities who, initially, have very limited knowledge about each other and their domain. The activity of the community of agents is shown in Figure 1. In this paper, the phrases ``solving a problem'' and ``problem-solving episode'' both refer to a single pass through these four steps, i.e. problem-solving is defined in terms of successfully executing actions (operators), not constructing plans. The performance of the community can be measured over successive problem-solving episodes in order to quantify the impact of collective memory. One measure of performance that we track is the number of times the agents must loop through step 3, which we call `rounds.'

The focus of this paper is on the conversion of run-time activity into cooperative procedures, which takes place in step 4 above and is detailed fully in Section 4. The agents can use these procedures in lieu of plans generated from scratch during step 3a. All agents presently work on the entire set of community goals, but future research will investigate the benefits of using collective memory to more effectively allocate the goals in step 2.