Information Technology Reference
In-Depth Information
cumstances in which it is favourable to change connectivity. The market used for these
simulations is very simple, it is not designed to reflex the intricacies of any particular
distributed market in particular. Instead it is designed to provide general insight into
the valuation of information in segmented markets. The results found within this paper
could be applied to any markets where information cannot flow freely. This includes
retail markets, OTC markets, and many others.
This model will differ from previous work in that it will model the micro-behaviour
of the traders. In both of the previous studies mentioned above [8,9], trade between
agents was abstract. When two agents were chosen to trade, their utility functions were
examined and an allocation of resources was calculated such that the utilities of both
agents were increased. In this study, we will use a well-established trading agent al-
gorithm to investigate the effect of the market constraints on the ability of agents to
identify the optimum price. In addition, an attempt will be made to modify the trading
agent algorithm to better cope with, or exploit, this situation.
2M thod
This section will first describe the structure and function of the markets that will be
investigated, before detailing the trading agents that will populate them.
2.1 Network Generation
Trading networks were constructed in which nodes represented traders and edges repre-
sented bi-directional communication channels. There are many possible network con-
figurations which could be investigated for their effect on market performance, includ-
ing lattices, Erd os-Renyi random graphs, small worlds, and graphs resulting from pref-
erential attachment. This paper will focus on the latter class of networks since they
exhibit some interesting properties, including the presence of well-connected “hubs”,
which have an intuitive appeal in terms of real world markets, where it would be ex-
pected that certain major investment banks would be much better connected than indi-
vidual investors.
An existing preferential attachment scheme is employed here [10]. A network of
N
unconnected nodes is gradually populated with
Nm
edges. In random order, each node
is consulted, and allocated an edge linking it to a second node chosen according to prob-
abilities calculated as
p
i
=(
n
i
+
δ
)
P
. Here,
P
is the exponent of preferential attachment
and remains constant,
n
is node's current degree (number of edges), and
δ
is a small con-
stant (0.1 for all results reported here) that ensures unconnected nodes have a non-zero
probability of gaining a neighbour. Self-connections and multiple connections between
the same pair of nodes were not allowed. All probabilities,
p
i
, were updated after ev-
ery edge was added. After
m
cycles through the population, the network was complete.
Note that every node will have a minimum of
m
edges, and a maximum of
N
1.
Markets explored here have a relatively high preferential exponent of
P
=1
.
5 in
order to generate networks that display a wide range of degrees. For all results reported
here,
m
=10. Initial tests showed that if
m
was significantly less than this value, the
market failed to converge as few agents were able to trade with their limited number of
neighbours.
−
