Figure 11 shows that the use of procedural memory in MW reduces the amount of runtime communication. As individual actors get better at guessing what kinds of joint activities work and how they will unfold, the amount of runtime communication needed to achieve their joint goals is decreasing.
This reflects possibly two things: the number of requests is decreasing and/or the percentage of agreed-to requests are increasing. The number of requests is decreasing either because actors are making fewer requests that are rejected, or because the actors are able to anticipate a request or action on the part of another actor without having to directly communicate. Our experimental data shows that by the tenth problem solving episode, with the use of procedural memory, the number of requests from one actor to another has decreased but the percentage of agreed-to requests has increased, i.e.,
Requests | Percentage agreed-to | |
Initial behavior | 19.6 | 73.7 |
Baseline system | 17.4 | 62.7 |
Learning conventions | 8.7 | 77.9 |