Decision making by intelligent agents: logical argument, probabilistic inference and the maintenance of beliefs and acts John Fox Advanced Computation Laboratory, Cancer Research UK (Previously ICRF), London WC2A 3PX, UK Peter McBurney Department of Computer Science Liverpool University Liverpool L69 72F Abstract PROforma is a language and a technology for designing, implementing, testing and delivering software agents that can operate in dynamic and uncertain environments. The agent specification language is based on R 2 L, a logic language that formalises reasoning, decision- making and plan execution using a combination of classical and non-classical logics. PROforma can be viewed as an object-oriented layer on top of R 2 L that reifies the logic into a small set of “tasks”, notably decisions, plans and I/O tasks. The heart of decision-making in both PRO forma and R 2 L is a logical decision model based on the creation and evaluation of arguments for and against alternative beliefs and actions. In previous papers we have provided proof-theoretic and model-theoretic semantics for a Logic of Argument and suggested that argumentation is a general framework which subsumes many different forms of uncertain reasoning as special cases, including qualitative and non-monotonic logics; quantitative representations like probability and possibility, and “linguistic” representations of belief. Current implementations support a simple monotonic interpretation of argument based decision-making but do not address requirements for defeasibility or probabilistic belief revision. In this paper we discuss how argumentation can provide a framework for integrating these approaches into a unified agent model. Introduction PROforma is an executable specification language for modelling decisions, plans and other tasks that an agent needs to have in its repertoire in order to achieve its goals (Fox and Das, 2000). The language is a versatile language for designing agent applications and has been successfully used in a variety of medical applications (e.g. Fox and Thomson, 1998). Software is available for authoring, verifying and testing applications in the language and enacting the specifications as autonomous processes, or as agents supporting human users. The PROforma language is defined around four classes of task: An enquiry is any process that acquires information during the enactment of a plan, whether by a dialogue with another agent, or through a sensor or other device. An action is any act carried out on the agent’s environment, whether a direct call to an effector or a request for action to another agent. A decision is a general procedure for making choices where there is uncertainty about the options, reliability of evidence, soundness of conclusions etc. A plan is a collection of tasks for achieving agent’s goals. Plan tasks can be To be presented at: Uncertainty Frameworks in Non- monotonic Reasoning. 9th Intl. Workshop on Non- Monotonic Reasoning (NMR'2002) Toulouse, April 2002.