In order to demonstrate the viability of the SMS with learning agents
and under real-world heavy usage conditions, we ran several tests on
the scenario shown in Figure 1. We implemented UIAs
that periodically (every 16 seconds) buy a query from some QPA using
the protocol described is Section 2. We also used
one Auction, AMA, and SCA, along with several UIAs and QPAs. All the
agents were deployed in machines all over our network. The UIAs kept
track of how long it took for the QPA's reply to arrive and used this
value in their learning. In general, the UIAs preferred fast and cheap
service, and they were willing to pay more for faster service. A QPA's
only preference was to increase its immediate profit, i.e. place its
bid in order to maximize its expected profit (remember that failing to
get a sale means the QPA gets zero profit).
We gave the agents different learning abilities (see
[7]). 0-level agents used a simple form of reinforcement
learning on the prices/values received. 1-level agents actually tried
to model the other agents (as 0-level agents) and took actions based
on predictions from these models.
Jose M. Vidal
jmvidal@umich.edu
Tue Sep 30 14:35:40 EDT 1997