1-level QPAs take advantage of price fluctuations by keeping models of
the QPAs and UIAs and use these to make better predictions as to what
they should bid. The 1-level models, while computationally expensive,
allow QPAs to track the individual agents more closely, thereby
identifying when a UIA is willing to pay more than the going rate.
Previous research has shown that the advantages of 1-level models can
be correlated to the price volatility (see [7]).
However, this strategic thinking is only successful against 0-level
UIAs. When we tested the 1-level QPAs against the 1-level UIAs we
found the QPA's performance on par with other similar 0-level QPAs.
The 1-level QPA's advantage was eliminated only at the cost of making
the UIAs keep 1-level models. From the systems' perspective, these
computations are ``wasted'' effort in that they do not contribute in
any way to the task of servicing the user. Notice, however, that they
would only be needed when the 1-level QPAs participating in the
auction are offering significantly different services. In these cases,
the buyers will want to determine which one of the QPAs offers the
``best'' (in its own opinion) service. Once they have determined this
then, perhaps, a new auction could be started that would service only
these QPAs so that the UIAs could then go back to being simpler
0-level agents and avoid the heavy computational costs. The
incentive/financing mechanisms for starting new auctions are still
under study.
Jose M. Vidal
jmvidal@umich.edu
Tue Sep 30 14:35:40 EDT 1997