Information Technology Reference
In-Depth Information
Carmel and Markovitch [7] describe a game-player that tries to analyze and learn
the strategy of its opponent. They discuss the benefits of using a model of the opponent
strategy, and give an algorithm called M* (a generalization of the standard minimax al-
gorithm) that attempts to exploit the opponent strategy. M* assumes that the opponent's
search depth and evaluation function are known, which is not the case in TAC SCM.
Chajewska, Koller, and Ormoneit [8] describe a method for predicting the future
decisions of an agent based on its past decisions. They learn the agent's utility function
by observing its behavior. Their approach is based on the assumption that the agent
is a rational decision maker. According to decision theory, rational decision making
amounts to the maximization of the expected utility [9]. In TAC SCM, we cannot apply
these techniques because the behaviors of individual agents are not directly observable.
Sales strategies used in previous TAC SCM competitions have attempted to model the
probability of receiving an order for a given offer price, either by estimating the proba-
bility by linear interpolation from the minimum and maximum daily prices [10], or by
estimating the relationship between offer price and order probability with a linear cumu-
lative density function (CDF) [11], or by using a reverse CDF and factors such as quantity
and due date [12,13], or by letting other agents set the price and trying to follow [14].
All these methods fail to take into account market conditions that are not directly
observable. They are essentially regression models, and do not represent qualitative
differences in market conditions. Our method, in contrast, is able to detect and forecast
a broader range of market conditions.
Regression based approaches (including non-parametric variations) assume that the
functional form which defines the relationship between dependent and independent
variables has the same structure. However, as shown in Figure 1, these functional re-
lationships have a different structure for different regimes. Therefore, an approach that
does not assume a functional relationship maybe the best way to identify a regime.
Wellman et al. [15] demonstrate a method for predicting future customer demand
in the TAC SCM game environment, and use the predicted future demand to inform
agent behavior. Their approach is specific to the TAC SCM situation, since it depends
on knowing the formula by which customer demand is computed. Note that customer
demand is only one of the factors for characterizing the multi-dimensional regime pa-
rameter space.
7
Conclusions and Future Work
We have presented an approach to characterizing and predicting economic market con-
ditions in markets for durable goods. Our approach recognizes that different market
situations have qualitative differences that can be used to guide the strategic and tac-
tical behavior of an agent. Unlike regression-based methods that try to predict prices
directly from demand and other observable factors, our approach recognizes that prices
are also influenced by non-observable factors, such as the inventory positions of the
other agents. Unlike price-following methods, our approach promises to enable an agent
to anticipate and prepare for regime changes, for example by building up inventory in
anticipation of better prices in the future or by selling in anticipation of an upcoming
oversupply situation.
