INTERRAP, like Ferguson's TOURINGMACHINES, is a layered architecture, with each successive layer representing a higher level of abstraction than the one below it [Müller, 1994][Müller et al., 1994][Müller and Pischel, 1994]. In INTERRAP, these layers are further subdivided into two vertical layers: one containing layers of knowledge bases, the other containing various control components, that interact with the knowledge bases at their level. At the lowest level is the world interface control component, and the corresponding world model knowledge base. The world interface component, as its name suggests, manages the interface between the agent and its environment, and thus deals with acting, communicating, and perception.
Above the world interface component is the behaviour-based component. The purpose of this component is to implement and control the basic reactive capability of the agent. This component manipulates a set of patterns of behaviour (PoB). A PoB is a structure containing a pre-condition that defines when the PoB is to be activated, various conditions that define the circumstances under which the PoB is to be considered to have succeeded or failed, a post-condition (a la STRIPS [Fikes and Nilsson, 1971]), and an executable body, that defines what action should be performed if the PoB is executed. (The action may be a primitive, resulting in a call on the agent's world interface, or may involve calling on a higher-level layer to generate a plan.)
Above the behaviour-based component in INTERRAP is the plan-based component. This component contains a planner that is able to generate single-agent plans in response to requests from the behaviour-based component. The knowledge-base at this layer contains a set of plans, including a plan library. The highest layer in INTERRAP is the cooperation component. This component is able to generate joint plans, that satisfy the goals of a number of agents, by elaborating plans selected from a plan library. These plans are generated in response to requests from the plan-based component.
Control in INTERRAP is both data- and goal-driven. Perceptual input is managed by the world-interface, and typically results in a change to the world model. As a result of changes to the world model, various patterns of behaviour may be activated, dropped, or executed. As a result of PoB execution, the plan-based component and cooperation component may be asked to generate plans and joint plans respectively, in order to achieve the goals of the agent. This ultimately results in primitive actions and messages being generated by the world interface.