Next: 5.7 Effect of optimizations Up: 5. Experimental Analysis Previous: 5.5 Planning

  
5.6 How is Memory for Prior Coordinated Activity Improving Performance?

The analysis presented so far confirms that the participants are getting better at coordinating their behavior when learning conventions. Several measures of external behavior reflect improvement. Overall communication and action effort are reduced; actors attempt and succeed at finding the optimal solution for moving an individual box more frequently; and the number of requests between actors is decreasing, while the percentage of agreed-to requests is increasing.

Why is the external behavior of the community improving? Our results show that learning conventions improves performance by predisposing actors to approach the problems in related ways.

We track conversations in order to measure when two actors are assessing a situation in related ways. One indication the two actors are assessing the situation in similar manners is that the listener is already working on the same goals as the requester. A stronger indication is when the listener is both working on the same goals and the listener's plan already includes a coordination point corresponding to the request. More technically, the listener's plan is considered related to the incoming request if any of the following are true:

1.
The incoming request is a request for a service to be provided and the plan contains either:
(a)
Communication about the same request; or
(b)
A wait operator for the same request.
2.
The incoming request is a request for a joint action and the plan contains either:
(a)
Communication about the same action; or
(b)
A wait operator for the same action; or
(c)
The same action.

The ideal situation would be that when one actor makes a request of another, the two actors always are working on the same goals and have plans containing the same coordination points.

Figure 13 measures how frequently two conversing actors have similar assessments of the situation. The left hand side of the figure measures the performance for the baseline system. For about half of the requests, the listener has the same goals and 30% of the time she has the same goals and a related plan. The right hand side of the figure measures performance when the actors use their memory of prior coordinated behavior to inform their activity. Here we see improvements in both statistics. More significantly, the curves are converging. In other words, the actors are learning to assess the same situation in a similar manner by extracting plans with related points of coordination to accomplish the same goals.

  

[baseline system doesn't improve ...] [learning system does]
Figure 13: Measuring how conventions lead to related approaches.

Despite the fact that actors are converging on coordination points, they do not converge on an identical mental structure for representing conventional behavior. We measured the overlap in the mental structures of independent actors in two ways. One way compares pairs of casebase entries from different actors to see how frequently they have the same plan. This entails having the same number of actions, each of which has the same action description (modulo the internal variable names). Because of their differing operator sets, lifters and hand-truck operators never the same plan. Even two lifters do not converge on the same plan: only 2.4% of their memories overlap. A stricter measure that requires that the entries are stored under the same contextual indices, never breaks 1.2%.


Next: 5.7 Effect of optimizations Up: 5. Experimental Analysis Previous: 5.5 Planning
Last Update: March 10, 1999 by Andy Garland