Using the hotel price prediction module as described in
Section 4, coupled with a model of its own
effect on the economy, *ATTac-2001* is equipped to determine its bids for
hotel rooms.

Every minute, for each hotel auction that is still open, *ATTac-2001* assumes that auction will close next and computes the marginal
value of that hotel room given the predicted closing prices of the
other rooms. If the auction does not close next, then it assumes that
it will have a chance to revise its bids. Since these predicted
prices are represented as distributions of possible future prices,
*ATTac-2001* samples from these distributions and averages the marginal
values to obtain an expected marginal value. Using the full minute
between closing times for computation (or 30 seconds if there are
still flights to consider too), *ATTac-2001* divides the available time
among the different open hotel auctions and generates as many price
samples as possible for each hotel room. In the end, *ATTac-2001* bids the
expected marginal values for each of the rooms.

The algorithm is described precisely and with explanation in Table 7.

One additional complication regarding hotel auctions is that, contrary
to one of our assumptions in Section 2.2
(Assumption 4), bids are not fully retractable: they
can only be changed to $1 above the current ask price. In the case
that there are current active bids for goods that *ATTac-2001* no longer
wants that are less than $1 above the current ask price, it may be
advantageous to refrain from changing the bid in the hopes that the
ask price will surpass them: that is, the current bid may have a
higher expected value than the best possible new bid. To address this
issue, *ATTac-2001* samples from the learned price distribution to find the
average expected values of the current and potential bids, and only
enters a new bid in the case that the potential bid is better.