Comparative Analysis of Frameworks for Knowledge-Intensive Intelligent Agents
Jones, Randolph M., Wray, Robert E., AI Magazine
One frequently taken approach toward b achieving human-level intelligent systems is to create foundational software systems that tightly integrate some number of representations and processes deemed sufficient for generating automated intelligent behavior. The design of these foundational software systems, which include both cognitive and agent architectures, have generally been based on some small set of theoretical principles. The agent architecture is an attempt to foster the development of uniform approaches for building intelligent systems. However, large-scale integrated software systems that attempt to approach human levels of intelligence through agent architectures exhibit some core commonalities across different architectures. For example, no matter the chosen architecture, there is a necessity for such systems to encode vast amounts of knowledge in efficient, organized, and maintainable ways. Additionally, these knowledge requirements have had relatively uniform effects on the evolution of these architectures, such that we observe a convergence of essential representations and processes across agent architectures.
A variety of frameworks currently exist for designing human-level intelligent agents and behavior models. Although they have different emphases, each of these frameworks provides coherent, high-level views of intelligent agency. However, more pragmatically, much of the complexity of building intelligent agents occurs in the low-level details, especially when building agents that exhibit high degrees of competence while interacting in complex environments. To highlight the emphasis of our observations about the knowledge necessary for human-level artificial intelligence, we call such agents knowledge-intensive agents. This term is also meant to distinguish such agents from smaller-scale, single-task agents (for example, service brokers) that are often fielded in multi-agent systems. Examples of fielded knowledge-intensive agents include a real-time fault diagnosis system on the Space Shuttle (Georgeff and Ingrand 1990) and a real-time model of combat pilots (Jones, Laird, and Nielsen 1999). Knowledge-intensive agents are also often used in "long-life" situations, where a particular agent needs to behave appropriately and maintain awareness of its environment for a long period of time (hours to days) while performing many different activities during the span of its existence. Additionally, knowledge-intensive agents must be engineered such that their knowledge can be easily modified (possibly by both extrinsic and intrinsic processes) as environment and task requirements change during deployment.
Transfer and generalization of results from one framework to others is usually slow and limited. The reasons for such limited transfer include differences in nomenclature and methodology that make it more difficult to understand and apply results, and the necessity of specifying low-level details that are not prescribed by the frameworks but that become important in actual implementation. In addition, high-level agent frameworks do not usually guide the agent developer in many finer-grained implementation issues, meaning that the frameworks underspecify necessary principles to build and field working intelligent agents. Our goal is to develop techniques that will minimize framework-specific descriptions and that bridge the gap between a framework's theory and the details of its implementation, especially clarifying which details are intrinsic to particular approaches and which are not. In the long run, this effort should foster reuse of architectural components and idioms across architectures as well as across individual agent models that use a single architecture.
This article reviews four existing agent frameworks in order to explore what they specify (and do not) about an agent's design and construction. The chosen frameworks have proven successful for building knowledge-intensive agents of various levels of complexity, or specifically address constraints on agents with high levels of competence (such as human behavior models). …