A Canonical Agent Model for Healthcare Applications
IEEE Intelligent Systems
A lthough the utility of autonomous agents and multiagent systems in healthcare applications is now well established, agent technologies themselves remain somewhat immature and, from a theoretical point of view, often ad hoc. We've developed a standard, or canonical, agent model that's intended to be both theoretically well motivated and technically well defined while permitting alternative instantiations. Our starting point is the domino agent model. 1 The Advanced Computation Laboratory
... on Laboratory developed this model for healthcare applications, but its design supports general-purpose cognitive agents. The domino model is similar to the classical beliefsdesires-intentions framework, but goes beyond BDI by defining a complete set of processes to transform between mental states, including a flexible decisionmaking framework based on logical argumentation. The domino model reflects current trends in software agent design, but it has broader justification in its embodiment of features common to many disciplines and theories of cognitive systems, including neuroscience and cognitive psychology as well as AI. 2 For example, a number of basic cognitive functions are widely held to be required by any intelligent agent: perception and interpretation of the agent's environment, goal setting and maintenance, problem solving, decision making, plan assembly, plan execution, and action selection. There is also a consensus on the types of representation on which these processes operate: beliefs, goals, and plans are assumed in a wide range of approaches. In general terms, these functions and representations are present across many theoretical approaches to cognitive agents.  The domino agent also provided the foundation for a practical agent language, ProForma (www. openclinical.org/gmm_proforma.html). 6 ProFormahas been used extensively to construct healthcare applications such as decision support and clinicalworkflow management. The domino model has been significantly extended within the Argumentation Services Platform with Integrated Components project. Funded by the European Commission, ASPIC involves a broad consortium of partners concerned with the uses of argumentation in agent systems, including nonmonotonic reasoning, decision making, interagent dialogue, and learning. Our canonical model captures the extended model in a general, implementation-independent way that provides a practical foundation for specific system implementations and agent-implementation tools. To address the need for canonical abstraction, we've adopted software engineering's concept of signatures-a technique for defining software patterns or invariant procedural properties. Healthcare applications have a number of additional requirements beyond the basic functions and representations that are common to many cognitive-system theories. (The "Related Work in Multiagent Healthcare Systems" sidebar describes four multiagent healthcare systems that, in different ways, illustrate these requirements.) On the basis of our experience with healthcare systems, we've identified three key requirements over and above the basic domino model: A communication capability for interactions between agents, which is important for multiagent A general autonomous agent system architecture has emerged from the healthcare application domain. Twelve invariant inputoutput patterns, or signatures, summarize its component properties, and shared data structures define the component interactions.