The figure above shows the main components of TacTex and their interaction.
Tasks are divided between a Supply Manager module and a Demand Manager module.
The Supply Manager handles all planning related to component
inventories and purchases, and requires no information about computer production except
for a projection of future component use, which is provided by the Demand Manager.
The Demand Manager, in turn, handles all planning related to computer sales and production.
The only information about components required by the Demand Manager is a projection of the current inventory and future component deliveries, along with an
estimated replacement cost for each component used. This information is provided by the Supply Manager.
The tasks to be performed by these two managers can bee seen as optimization tasks: the Supply Manager tries to minimize the cost of obtaining the components
required by the Demand Manager, while the Demand Manager seeks to maximize the profits from computer sales subject to the information provided by the
Supply Manager. In order to perform these tasks, the two managers need to be able to make predictions about the results of their actions and the future
of the economy. TacTex uses three predictive models to assist the managers with these predictions:
a Supplier Model, a Demand Model, and an Offer Acceptance Predictor.
The Supplier Model keeps track of all information available about each supplier, such as outstanding orders and the component prices that have been offered
in the past.
Using this information, the Supplier Model can assist the Supply Manager by making predictions concerning future component availability and prices.
The Demand Model tracks the customer demand in each of the three market segments, and tries to estimate the
underlying demand parameters in
each segment. With these estimates, it is possible to predict the number of computers that will be requested on any future day.
The Demand Manager can then use these predictions to plan for future production.
When deciding what offers to make in response to customer requests, the
Demand Manager needs to be able to estimate the probability of a
particular offer being accepted (which depends on the prices offered by the other agents).
These predictions are handled by the Offer Acceptance Predictor.
Based on past bidding results, the Offer Acceptance Predictor produces a function for each request that maps offer prices to the predicted probability of
winning the order.
When competing in a number of games against a fixed set of opponents, such as during the competition finals,
TacTex adapts its early and late game strategies in response to the results of past games.
Depending on the orders from other agents, component prices on the first day of the game may be much higher or lower than they will be later in the game.
TacTex analyzes past games to determine the exact set of requests it should send on the first day to get the lowest prices.
Results from past games are also used to determine the probabilities of having offers to customers accepted during the final few game days, when opponents tend to either have a surplus or shortage of computers available to sell.