Information Technology Reference
In-Depth Information
73,44
70,41
63,74
58,70
57,15
43,70
Figure 11.14 Experiment 7: real time, introducing the first trustworthy strategy (measuring Credits).
(Reproduced with kind permission of Springer Science+Business Media C
2005)
higher cost; so the agent who maximizes trust is penalized with respect to gain. The agent who
maximizes following the utility function chooses agents with less ability and willingness; for
this reason it performs fewer tasks than the other agents. 11
EXP8 (see Figure 11.16) was performed for 200 tasks and 250 simulations. The results
refer (1) to the tasks performed; 12 (2) to the gains; (3) to time spent. The policies are: random ;
cost trust (that uses the utility function); trust ambient (that uses trust as usual); low cost (that
always delegates to the less expensive); first trust (that delegates to the first over a certain trust
threshold), stat (that performs statistical learning).
The most important result is about gains; the agent who explicitly maximizes them has a
great advantage; the trust strategies perform better than the statistical one, even if they all do
not use information about gains but about tasks achieved. It is even interesting to notice that
the worst strategy results delegating always to the least expensive: in this case costs are related
to performance (even if in an indirect way) and this means that cheap agents are normally bad.
Delegation and Monitoring
In order to model a more complex contract net scenario, we have introduced the possibility of
the delegator monitoring the performance of the delegated agents into the intermediate steps
11 We have chosen a little difference between costs and gains in order to keep costs significant: setting lower costs
(e.g. 1-10) makes them irrelevant; and the trust agent performs better.
12 Multiplied 10 in order to show them more clearly.
 
Search WWH ::




Custom Search