- ...Brooks, Durfee
- Artificial Intelligence
Laboratory, University of Michigan, Ann Arbor, MI 48109,
{chbrooks,durfee}@umich.edu
.
.
.
.
.
.
.
.
.
.
- ...Fay
- Department of Economics, University of
Michigan, Ann Arbor, MI 48109, scottfay@umich.edu
.
.
.
.
.
.
.
.
.
.
- ...Das, Kephart
- Institute For Advanced Commerce, IBM
Research, PO Box 704, Yorktown Heights, NY 10598, {rajarshi,kephart}@watson.ibm.com
.
.
.
.
.
.
.
.
.
.
- ...MacKie-Mason
- School of Information and
Department of Economics, University of Michigan, Ann Arbor, MI 48109,
jmm@umich.edu
.
.
.
.
.
.
.
.
.
.
- ...
- This work was supported in part by an IBM University
Partnership Grant, DARPA grant F30602-97-1-0228, part of the
Information Survivability program, and by the National Science
Foundation under grant IIS-9872057
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
- ...consumer.
- We have calculated the exact
solution for two of our price schedules and found that the continuous
approximation was not very good for a market with ten articles, but
was quite close for N=100 articles.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
- ...preferences.
- We plan to explore model-based
learning in subsequent work. First we will need to make the
environment more complex. Our stationary environment with linear
preferences permits a smart producer to learn everything in just one
transaction period if the number of articles consumed is observable.
Of course, such powerful learning possibilities are not realistic.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
- ...13#13's.
- In Section 3 we use simulation methods to analyze a
model with incomplete information, in which the producer does not
know 14#14, or any individual's 13#13.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
- ...here.
- The details of the
derivations are available from the authors on request.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
- ...welfare.
- We include
results for block pricing, an intermediate case with three
parameters, though we did not include the derivation above to
conserve space.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
- ...schedules
- Prices are discretized in this estimate so as to
approximate a ``degree of completion.''
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
- ...complete
- Recall that the plots show results averaged over
several runs with randomized starting values. The global peak of the
mixed bundling landscape was found less often than that of two-part
tariff.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
- ...dominate
- The unconstrained nonlinear schedule has N = 100
free parameters now; amoeba was not able to search the resulting
100-dimensional space in an acceptable amount of time.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
- ...Hill''.
- For
example, each time a competitor changes its price
schedule and attracts a different set of consumers, or consumers
switch between producers to do their own exploration, the effective
population of consumers (and thus the distribution of preferences and
the resulting profit landscape) facing the first producer will change.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.