Information Technology Reference
In-Depth Information
in the slower initial convergence seen in the most well-connected agents. One might
imagine that since these well-connected, but initially misguided individuals have a very
significant influence on their neighbours, they might retard the market's convergence as
a whole. Results suggest that this does not occur for the topologies considered here, but
one could imagine market structures in which the hubs are so large and scarce that they
could disturb the market for some time. As it stands, the simple adaptive learning rate
rule could obviously be modified to better suit the more well-connected individuals, or,
alternatively, separate rules could be used.
The adaptive learning rate rule used in this paper is very simple. It was chosen in
order to demonstrate that the quality of an information source could be an important
factor in making trading decisions, and that this could be used by trading agents to
improve their valuations. It is not an attempt to provide an optimal rule. Their are many
other possible factors which could be incorporated in order to make this rule more
sophisticated.
As it currently stands this rule has several weaknesses. The most obvious is that it
relies on information which may not be publicly known. First, it uses the maximum con-
nectivity ratio present within the market in order to normalise the rate of change. This
is necessary in order to ensure that learning rates were scaled to fall within the same
range as that employed by standard ZIP agents. Second, whenever an agent adapts, it
uses the connectivity of the shouter in order to determine how much attention should
be paid to the shout. In real situations it seems unlikely that an agent would have access
to either of these kinds of information. In order for an agent to know the maximum
connectivity ratio, it would be necessary for the agent to know how the whole of the
market was structured. For the simulations reported here, this information is easy to
obtain. However, in real markets it seems highly unlikely that this information would
be available. The only probable way for an agent to know the whole market structure
is to be in contact with every agent in the market. If this were possible, then it is likely
that other agents would also be completely connected, and an individual agent's con-
nectivity would cease to be an issue. The connectivity of an individual trader may be
even more difficult to obtain. In the studies reported here, every agent knows its own
connectivity and that of its neighbours. In reality it is unlikely that the agent would pos-
sess this second piece of information, or perhaps even the first. The only realistic way
to determine how well-connected another agent is, is to obtain the information from
the agent directly. As has been suggested, however, it seems unlikely that agents would
want to give this information away, as our results suggest that it is valuable.
This begs the question, if it is impossible to obtain the necessary connectivity in-
formation in reality, what practical significance do the results presented here have?
Although it may be impossible to obtain exact connectivity information, it is not im-
possible to generate an evaluation of the information quality of ones trading partners
in a real market. Moreover, information quality can be measured via indicators other
than connectivity. In the case of human markets, traders may implicitly evaluate many
aspects of their trading partners when deciding the significance of a piece of informa-
tion. These aspects may include estimates of the partner's size (is it an investment bank
with many traders or is it an sole trader), reputation, market position, market experi-
ence, etc. As yet, it may be difficult for an artificial trader operating in a real market,
