4 A Model of Co-evolving Social Agents

4.2 The Results

In figure 9 and figure 10 the attendance patterns of the agents during the eight runs are displayed. The most obvious feature is the difference between the patterns under the crowd-avoiding and friendly runs; under the crowd-avoiding scheme attendance appears far more stochastic compared to those under the friendly scheme where there is obvious coordination. This is unsurprising given that the crowd-avoiding utility scheme encourages the competitive discoordination of behaviour whilst there is a considerable advantage to (at least somewhat) coordinating action with one's `friends' under the friendly scheme.

Figure 9. Attendances for the four runs under the crowd-avoiding scheme
(grey=went, black=stayed at home)

Figure 10. Attendances for the four runs under the friendly scheme
(grey=went, black=stayed at home)

The first run exhibits the least regularity - it looks like the output from a stochastic process*1. It appears that while listening and the friendly utility scheme encourage the emergence of heterogeneity among agents (i.e. there is a differentiation of strategies), imitation encourages a similarity of behaviour between agents (apparent in the vertical stripes in the ca+i run and the uniformity of the fr+i run).

In table 1 and table 2, the average utility gained over the last 30 weeks and over all agents is shown for each run of the simulation. The utility gained under the crowd-avoiding and friendly cannot be directly compared. Under the crowd-avoiding scheme (table 1) the only significant difference (at the 5% level) is between the ca and ca+li runs. In the runs using the friendly scheme (table 2) the only significant difference (this time at the 1% level) is between the fr+i run and the others.

The next figures (figures: 11, 12, 13, 14, 15, 16, 17, and 18), show some of the specific causation between the talk and action expressions of the ten agents. To keep the diagrams manageable I have limited these to the last three weeks of each run of the simulation. These figures only show the causation due to the saidBy, saidByLast and wentLastWeek primitives that are active (where by `active' I mean a saidBy or saidByLast primitive in a simulation where listening is enabled and where it isn't logically redundant). So they do not show any causation via attendance statistics (e.g. averageOverLast, numWentLast), or the self-referential primitives (e.g. ISaidYesterday, IPredictedLastWeek and IWentLastWeek) since I wish to focus on the embedding and these are not so relevant to this concern*2. In these figures there is a small box for the talk and action expression of each agent (numbered upwards from 1 to 10) - so, for example, that the topmost box under the `A' label represents the action strategy of barGoer-10. The numbers in the boxes are the total number of backward causal lines connected to that box if one followed the causation backward (restricted to the last three weeks only). This number is thus an indication of how socially embedded the agent is at any point in time - a larger number indicates that there is quite a complex causal chain determining the action (or communication) of that agent, passing through many other agents. A detailed example of this (barGoer-6 at the end of the ca+l run) is analysed in greater detail below.

Figure 11. Causation net for run under crowd-avoiding scheme with neither listening nor imitation enabled (ca)

Figure 12. Causation net for run under crowd-avoiding scheme with only listening enabled (ca+l)

Figure 13. Causation net for run under crowd-avoiding scheme with only imitation enabled (ca+i)

Figure 14. Causation net for run under crowd-avoiding scheme with both listening and imitation enabled (ca+li)

Figure 15. Causation net for run under friendly scheme with neither listening nor imitation enabled (fr)

Figure 16. Causation net for run under friendly scheme with only listening enabled (fr+l)

Figure 17. Causation net for run under friendly scheme with only imitation enabled (fr+i)

Figure 18. Causation net for run under friendly scheme with both listening and imitation enabled (fr+li)

To enable a comparison of the general levels of embedding I have tabulated the average of the last two weeks of the total of these indicators over all the agents. These numbers are shown in table 3, and table 4. These are indicative only - they merely suggest that the crowd-avoiding runs of the simulation with listening enabled are more embedded that any of the other runs, with the crowd-avoiding run with listening only enabled, the most. In order to check these a more detailed study of the causation involved is required.

Generated with CERN WebMaker