Multi-Agent Architecture for Knowledge-Driven Decision Support - Distributed Artificial Intelligence, Agent Technology, and Collaborative Applications - page 31

• Collaborative in their ability to interact with other agents in exhibiting the goal-oriented behav-

ior

• Adaptive in their ability to learn with experience

Agent-based systems may consist of a single agent engaged in autonomous goal-oriented behavior,
or multiple agents that work together to exhibit granular as well as overall goal directed behavior. In the
general multi-agent system, the interoperation of separately developed and self-interested agents pro-
vides services beyond the capability of any single agent model. Mutli-agent systems provide a powerful
abstraction that can be used to model systems where multiple entities, exhibiting self directed behaviors
must coexist in a environment and achieve the system wide objective of the environment.
Intelligent Agents are action-oriented abstractions in electronic systems entrusted to carry out vari-
ous generic and specific goal-oriented actions on behalf of users. The agent abstraction manifests itself
in the system as a representation of the user and performs necessary tasks on behalf of the user. This
role may involve taking directions from the user on a need basis and advising and informing the user of
alternatives and consequences (Whinston, 1997). The agent paradigm can support a range of decision
making activity including information retrieval, generation of alternatives, preference order ranking of
options and alternatives and supporting analysis of the alternative-goal relationships. In this respect,
intelligent agents have come a long way from being digital scourers and static filters of information to
active partners in information processing tasks. Such a shift has significant design implications on the
abstractions used to model information systems, objects or agents, and on the architecture of informa-
tion resources that are available to entities involved in the electronic system. Another implication is
that knowledge must be available in formats that are conducive to its representation and manipulation
by software applications, including software agents.

decision Trees and IdSS

Models of decision problem domains provide analytical support to the decision maker, enhance under-
standing of the problem domain and allow the decision maker to assess the utility of alternative deci-
sion paths for the decision problem (Goul and Corral, 2005). Decision Trees are a popular modeling
technique with wide applicability to a variety of business problems (Sung et al., 1999). The performance
of a particular method in modeling human decisions is dependent on the conformance of the method
with the decision makers' mental model of the decision problem (Kim et al., 1997). Simplicity of model
representation is particularly relevant if the discovered explicit models are to be internalized by decision
makers (Mao and Benbasat, 2000) Decision Trees represent a natural choice for IDSS whose goal is to
generate decision paths that are easy to understand, explain and convert to natural language (Sung, et
al., 1999). The choice of decision trees as the modeling methodology affords the ability to incorporate
inductive learning in the IDSS. Decision trees are among the most commonly used inductive learning
techniques used to learn patterns from data (Kudoh and Haraguchi, 2003; Takimoto and Maruoka, 2003).
The ID3, C4.5 and SEE5 algorithms provide a formal method to create and model decision rules from
categorical and continuous data (Takimoto and Maruoka, 2003; Kudoh and Haraguchi, 2003). Kiang
(2003) compared multiple machine learning techniques and found that the decision tree technique had
the most interpretive power. They suggest the use of multiple methods in systems for effective intel-
ligent decision support.