Information Technology Reference
In-Depth Information
If
R
(
H
t
−
1
)
R
(
H
t
)
•
(A) Desired
≤
If
P
∗
(
H
t
−
1
,
Q
i
)
<
p
t
•
(B1) Occurred due to outdated information
(B2) Occurred due to a misleading non-anonymous quote If
R
(
H
t
i
)
<
R
(
H
t
−
1
)
•
If
R
(
H
t
)
<
R
(
H
t
−
1
)
•
(B3) Occurred due to a misleading anonymous quote
•
(C) Necessary due to auction dependencies
If
t
/
∈
A
,
B
1
,
B
2
,
B
3
,
C
Using the same data obtained for the experiments reported in Table 3, we measured
the percentage of transactions in (A), (B1), (B2), (B3), or (C) for the different quot-
ing mechanisms we tested with AON auctions. The results are shown in Table 4. Note
that the Full Schedule quote was the least misleading. Had we used a non-anonymous
full schedule quote for decision making, we would have entirely avoided (by defini-
tion) misleading quotes, potentially increasing the overall performance even further.
The marginal unit and standard quotes were similarly misleading, providing compa-
rable results regardless of quote anonymity. Thus, personalizing these two quotes for
decision making does not seem likely to help. Marginal unit quotes appeared to pro-
vide a relatively high percentage of desired trades (but poor overall results), which is
somewhat expected given such a highly conservative quote.
Ta b l e 4 .
Analysis of transactions in AON auctions. (*) Offline simulation only.
# Quoting (A) (B1) (B2)* (B3) (C)
1 Full schedule 62.1% 20.7% 0% 17.8% 16.1%
2 Marginal unit 64.8% 17.6% 26.2% 25.2% 3.5%
3
Standard
57.4% 22.8% 24.8% 23.1% 7.1%
5
Discussion and Future Work
Our study of the AON auction provides evidence for the viability and potential benefits
of accounting for indivisibility constraints in market-based allocation, without resorting
to fully combinatorial auction designs. Whether one should adopt an AON auction or
standard divisible auction depends on the specific setting. Relevant factors include:
1. Expressivity. Divisible bids allow expressing only convex valuations, whereas indi-
visible ones do not have such limitation. Would agents with nonconvex valuations
refrain from participating in auctions with mandatory divisibility?
2. Undesired trades. If agents with nonconvex valuations do participate in a divisible
auction, they risk loss-producing transactions. How much they suffer as a result
depends on the degree of nonconvexity.
3. Computation. Our incremental algorithms provide a relatively efficient way to op-
erate AON auctions, which should be fast enough for several practical applications.
Standard auctions, however, are still faster to operate, and more predictable since
performance is less dependent on the number of units offered for sale.
4. Quote communication. Our experiments showed that the level of detail in price-
quote information can play an important role in overall efficiency, and in particular
that simple marginal quotes were not enough for AON auctions to improve on the
