Development Research Map

Architectures

An agent architecture defines the components of an agent and its interrelationships. In the agent community three different kinds of architectures exist [WJ94]. It is interesting that none of the architectures is directly reflected by the standards (mainly because architectures considered, concentrate the agent internal concepts). Until now the standards and architectures seem to have a minimal overlap even though they should address the same topics.

Deliberative architectures are based on the physical-symbol system hypothesis [NS76], on the foundations of logic and on theorem provers. It is therfore assumed that an agent has a model of its environment (symbols) which can be used to deduce general intelligent action. Several different approaches exist, emphasizing different AI techniques as central agent "brain". Planning agents use a planner to construct sequences of actions to reach a certain goal. Main problem of this approach is the complexity and time consumption of plan construction [Cha87]. Mentalistic Agents represent an alternative approach is based on the notions of mental states (see Research Map MAS, internal concepts).

Reactive architectures result from the limitations imposed by symbolic AI. These types of agents do not have a complicated mind, but instead are constructed in a way that allows them to react to a changing environment by their "instincts". As the diction already indicates these architectures are associated with observations of behaviours from the animal world. E.g. an ant colony consists of different but very simple individuals but the colony itself exhibit more intelligent behaviour than one would expect from the ants seen in isolation (-> emergence). The subsumption architecture [Bro85] proposes a layered design of competing task-accomplishing behaviours. Lower layers exhibit more primitive kinds of behaviour, and have precedence over layers further up the hierarchy. A lot of other different reactive architectures exist (see [WJ94]). Most researchers agree that reactive agents are not well-suited for many kinds of problems.

Hybrid Architectures try to combine the advantages of the above mentioned paradigms with the aim of an integrated effective and efficient agent behaviour. Therefore AI components and reactive elements are subsumed into one design model. A well-known example is INTERRAP [Mue93], which consists of a layered world model and execution entity. In this architecture between purely reactive, planned and social behaviours is distinguished. Other approaches can be found in [WJ94].

[NS76]	A. Newell, H. A. Simon. Computer science as empirical enquiry. Communications of the ACM, 19:113–126, 1976.
[Cha87]	D. Chapman. Planning for conjunctive goals. Artificial Intelligence, 32, 1987.
[WJ94]	M. J. Wooldridge, N. R. Jennings. Agent Theories, Architectures, and Languages: A Survey. Workshop on Agent Theories, Architectures & Languages ECAI'94. Springer-Verlag. PP. 1-22. 1995.
[Woo01]	M. Wooldridge. An Introduction to MultiAgent Systems. John Wiley & Sons LDT, 2001.
[Bro91]	R. A. Brooks. Intelligence Without Reason. In Proceedings of the 12th International Joint Conference on Artificial Intelligence, IJCAI-91. Morgan Kaufmann publishers Inc. PP. 569-595. 1991.
[Mue93]	J. Muller, M. Pischel". The Agent Architecture InteRRaP: Concept and Application. Technical Report RR-93-26, DFKI Saarbrucken, 1993.