Information Technology Reference
In-Depth Information
(which are multiple English auctions modified with a deadline, proxy bidding and dis-
crete bids) bidding until one's valuation is no longer always the optimal strategy and in
Continuous Double Auctions (CDAs) (which are a symmetric auction mechanism with
multiple buyers and sellers) there is no known optimal strategy [7].
Given this background, there has been considerable research endeavour to develop
trading agents with heuristic strategies that are effective in particular marketplaces
[18,22]. Though more of a black art than an engineering endeavour at present, we
believe the design of successful strategies in such marketplaces can nevertheless be
viewed as adhering to a fundamental and systematic structure. To this end, in this pa-
per, we provide a general framework for designing strategies which is simple enough
to be applicable in a broad range of marketplaces, but modular enough to be used in
the design of complex strategic behaviour. We believe such a model is important for
the designers of trading agents because it provides a principled approach towards the
systematic engineering of such strategies which, in turn, can foster more reliable and
robust strategies.
As there is no systematic software engineering framework currently available for
designing strategies for trading agents, this paper advances the state of the art by pro-
viding the first steps towards such a model. Specifically, our framework is based upon
three main principles:
1. An agent requires information about itself and its environment in order to make
informed decisions.
2. An agent rarely has full information or sufficient computational resources to man-
age all the extracted information.
3. Given its limited computational resources and information, an agent needs to em-
ploy heuristics in order to formulate a successful strategy.
In order to operate in such situations, we advocate a multi-layered design frame-
work. We believe this is appropriate because most strategies can be viewed as breaking
down the task of bidding into a set of well defined sub-tasks (such as gathering rele-
vant information, processing that information and using that processed information in a
meaningful manner). This decomposition can be viewed as a series of (semi-) distinct
steps that are handled by different layers. Furthermore, our aim is to ensure our model
is sufficiently abstract to be used as the agent model in more general agent-oriented
software engineering frameworks, such as Gaia [23] and Agent UML [1]. To this end,
our framework is inspired by the distinction made in economics between information
and the knowledge derived thereof [6], and is augmented by the Behavioural Layer
(since the behaviour dictates which knowledge an agent seeks within an environment).
Specifically, our framework consists of three layers: the
Information
,
Knowledge
and
Behavioural
layers (hence we term our framework the
IKB
model hereafter).
In more detail, the information layer records raw data from the market environment.
This is then processed by the knowledge layer in order to provide the intelligent data
which is used by the behavioural layer to condition the agent's strategy. To illustrate
the use of our framework, we chose two example marketplaces that are popular for
trading agents. Firstly, we consider the marketplaces with one auction protocol, the
CDA, which is widely used in trading stocks. We place a number of the standard CDA
