Information Technology Reference
In-Depth Information
In short, the actions and sales performed by the seller and buyer agents are dynamic in nature, and
for optimization are best represented as a Dynamic Programming formulation.
In a typical application of this nature, the preferred system architecture is a
Blackboard
, with open
information about all offered trades in a given period available to all participating agents, for possible
matching individually or by consortium, for a particular commodity (Nag, 2007). The internal agent
architecture, preferably common to all participating agents, includes
domain knowledge
as an under-
standing of the trading task in terms of the tasked trading volume and the commodity to be traded.
Another part of the internal agent architecture is the
communication capability
that enables an agent to
process the community information versus the trading volumes and seek for help from other interested
agents in forming consortiums.
The implementation of an agent is an algorithmic software representation of a set of priorities and
working rules. The basic internal architecture is shown in Figure 1. A simplified set of rules and actions
is shown in Figure 2, in which, local and domain knowledge are both included. For reasons of simplicity,
the domain knowledge does not show details of trading parameters, such as product characteristics used
to find a trade, or knowledge of agent's objectives in making a trade. To implement the consortium, an
additional part of the domain knowledge of an agent is a map that has information about surrounding
agents. Without a supervising authority, the map information with each agent is incomplete, but infor-
mation overlaps exist between adjacent agents extending the available information.
In agent trading, the consortium effect is implemented as a combination of the map and message
passing. The algorithms for map determination and for messaging are shown in Figure 3. There are
two methods of message passing, broadcast and point-to-point. When an agent needs to find a potential
trading partner, it is best to broadcast trade requirements and conditions to many agents. To identify
Figure 2. Agentdomain algorithms
A
i
: represents agent
i
.
T(A
i
): represents a set of tasks of A
i
t
j
(A
i
): represents a sub-task
j
of A
i
I(A
i
): represents Input of A
i
O(A
i
): represents Output of A
i
P(I(A
i
)): priority of Input of A
i
P(O(A
i
)): priority of Output of A
i
D(T(A
i
)): set of domain for specific task of A
i
L(T(A
i
)): local knowledge of A
i
T(A
i
) = ∪
j
t
j
(A
i
)
{
Begin Trade}
Select Input (I(A
i
), P(I(A
i
)): select input according to the priority
Send Output(O(A
i
), P(O(A
i
)): send out output with a priority.
Apply (Select Input (I(A
i
), P(I(A
i
)), D(T(A
i
)), L(T(A
i
)), Send Output (O(A
i
), P(O(A
i
)))) : select input, apply domain knowl-
edge.
{
End Trade
}
Search WWH ::
Custom Search