Information Technology Reference
In-Depth Information
from the KL. If the trend suggests a decrease in price, the BL then queries the predicted
lowest ask price of the flight auction, and a bid is placed in that auction when that
minimum is reached, if such flight tickets are required in the optimal plan. Conversely, if
an increasing trend is predicted in a flight auction, we face a trade-off between acquiring
all the tickets in such an auction immediately at the current lowest price, and waiting in
case the agent does not manage to acquire the
scarce
hotel rooms required in the optimal
plan, which could make the flight tickets redundant (since they are no longer required in
the optimal plan and represent a loss). We implement the trade-off by spreading our bids
in a flight auction over the remaining length of the TAC game. For example, if 4 tickets
are required from a particular flight auction with an increasing trend, we could buy a
single ticket every minute over the next 4 minutes, rather that buying all 4 immediately.
Next, we have the 8 hotel auctions, with a random one clearing (and closing) every
minute. Thus, every minute, as the optimal plan changes, we update our bid in those
auctions that are yet to close. Now, there is uncertainty in being able to acquire all the
items required in the optimal plan, particularly at the beginning of the game. Further-
more the optimal plan typically changes during the game resulting in an item no longer
being required in the optimal plan as the game progresses. Thus, bidding too high ini-
tially does not pay off since such a bid could result in that item still being acquired.
Thus, our agent does not bid for a hotel room at its marginal profit (see Definition 7),
but rather bid low at the beginning of the game and gradually increases its bid for a
room towards its marginal profit as the game progresses, bidding its marginal cost after
the 7th minute before the last hotel auction closes.
Finally, we have the 12 entertainment auctions. Here, we use the RB strategy (see
Section 4) to bid in those CDAs. In particular, we have 12 RB traders that bid for the
items required in the optimal plan. The agent further instructs the RB trader to buy
cheap
in auctions that do not influence the optimal plan, and sell
high
all the items
that it holds, if the agent can thus be more profitable rather than using such items in its
optimal plan. We now consider the knowledge required for the bidding behaviour.
5.2
The Knowledge Layer
Here, we principally require the optimal plan which is given as the solution to an op-
timisation problem. The agent searches for the plan that maximises its profit, which
is the total utility of the packages less their estimated cost. The utility of a package
is determined by a client's preferences, which is queried from the IL. Furthermore, the
optimisation problem is constrained by different requirements of a feasible package, for
example a client needs to stay in the same hotel for the duration of his/her stay or the
client is required to stay in a hotel during the length of his/her stay [22], with additional
constraints imposed as hotel auctions close. We also consider the additional knowledge
of the predicted clearing price of the hotel auctions and of the flight auctions (based on
the trend of flight prices in those auctions) to estimate the cost of a plan.
Now, for the hotel auctions, we calculate the marginal profit of hotel rooms required
in the plan, to form the bidding price in the active hotel auctions. This is carried out
by considering the next best package if a particular hotel room in the optimal plan
cannot be acquired. The drop in profit then represents the marginal profit of that hotel
room. Next, for the flight auctions, the KL estimates the trend of the flight prices, by
