Multi-Agent System Research Map

Internal Artificial Intelligence Concepts

Agent architectures describe the means by which agents can reason about their agents in an environment. For intelligent behaviour it is necessary - in addition to the concepts for an internal agent structure - that special artificial intelligence mechanisms are utilized. On the one hand exist mechanisms that allow inferencing over knowledge which means that new facts can be deduced from the contained knowledge and on the other hand there are mechanisms that allow learning new things and integrate them into the agents knowledge.

Rule-based systems are a well-known approach for applying declarative knowledge and has its root in the expert system area. Such a system consists of a rule-base for storing the domain-specific rules and an inference engine that selects applicabe rules and fires them when on the knowledge base [RN03][Fri03]. One problem of rule-based systems is that they can only learn by adding new rules to the existing rule set what cannot easily be automated. It is a natural approach to try to combine the advantages of rule-based systems with agents, even though it is not desireable to base to complete functionality of an agent on rules. This is because rule systems are well suited for certain kinds of problems and should be seen as an valueable extension to the procedural approach, but have their limitations as well [Li91].

Case-based reasoning is a new research area that tries to overcome some limitations of the rule-based reasoning. It is an approach that utilizes the knows cases to solve new problems. Therefore for a new problem the nearest cases from the case library are extracted and the soluation that worked in the past is adapted and used in the new situation. When this solves the problem the case is stored in the case base, when it does not it is used to identify exceptions [AP94].

Other classical artificial intelligence mechanisms allow learning more directly. Neural nets [Smi03] are one well-known approach to train a model with a preparation set of input-output values and then use the model to calculate values for unknown inputs. A disadvantage of the neural net approach is that the reason for the produced output cannot be extracted. Therefore it is not possible to explain the produced result. Other known learning approaches include statistical learning, decision trees, inductive logic programming, reinforcement learning and explanation-based learning [Nil01].

It remains a research objective to find out, which learning mechanisms can be combined with the existing internal architectures and therefore lead to the conceptualization of intelligent agents.

[RN03]	S. Russel, P. Norvig. Artificial Intelligence: A Modern Approach, Second Edition, Prentice Hall, 2003.
[Fri03]	E. Friedmann-Hill. JESS in action: rule-based systems in Java, Greenwich, CT: Manning, 2003.
[Li91]	X. Li. What's so bad about rule-based programming? IEEE Software, 1991, Sept., 103-105.
[AP94]	Aamodt, A.; Plaza, E.: Case-Based Reasoning: Foundational Issues, Methodological Variations, and System Approaches, In AI Communications, 7(1), 1994.
[Smi03]	L. Smith. An Introduction to Neural Networks.
[Nil01]	N. Nilsson. Introduction to Machine Learning. Book draft. Not yet published.