Modelling Reactive Behaviour in Vertically Layered Agent Architectures

1 I n t r o d u c t i o n Over the past few years, several different architectures for autonomous systems have been proposed in the (D)AI literature (e.g. [2, 8, 3, 14, 4, 7, 10, 5, 16, 15]). An important class of approaches to modeling systems that have to behave in a goal-directed manner in a complex, changing environment are layered architectures. This approach regards an agent as consisting of several hierarchical functional modules, representing the different requirements on an agent, such as reactivity, efficiency, goal-directed behaviour, and coordination with others, as well as representing different qualities and levels of abstraction concerning the agent's knowledge (e.g. from raw sensor data to the description of complex, rather abstract situations). Basically all these approaches are somehow linking the input into an agent (its perception) into a kind of output from the agent (normally regarded as the actions the agent performs). Most of them define an agent cycle using perception in order to update the agent's internal state, i.e. its beliefs about the world, then use this world model as a basis to do some kind of decision-making (planning), possibly taking into account other agents, finally leading to a decision as to what to do next, i.e. to the actions to be performed in the next agent cycle. Possible layers of these agent models incorporate perception and action, reactivity (behaviour-based layer), local planning, cooperation, modeling, intentions, and learning. In this paper, we identify and describe several basic classes of layered architectures. We focus on a specific class, namely vertically layered architectures, and describe a concrete instance, the agent architecture INTERFtAP. INTERRAP consists of two basic units, the agent control unit and the agent knowledge base, which both share a hierarchical structure. The four control layers are (from lower to upper): the agents' world interface definition (WIF); the behaviour-based