weaknesses with respect to their relevance for military simulations. To model all the details of the complex behavior involved in the tank platoon's hasty defense described in the vignette of Chapter 2, an architecture would have to encompass all the phenomena addressed in all the following chapters: attention and multitasking, memory and learning, decision making, situation assessment, and planning, all subject to behavioral moderators. No existing architectures approach this all-encompassing capability. This does not make these architectures useless—many have been demonstrated to be quite useful despite their limitations—but it does suggest areas of improvement for each.
One research and development path is obvious: start with one of the existing architectures, and expand it in the areas in which it is weak. For example, an emerging ACT architecture, ACT-R/PM, adds EPIC-like perception and motor capabilities to the ACT-R architecture (Byrne and Anderson, 1998). Each of the reviewed architectures could be expanded in this way, as needed for task performance. This approach represents a time-honored path with almost guaranteed incremental payoff, but may eventually encounter boundaries as the architectures reach their limits of expressibility.
Another research and development path might prove more fruitful: combine the strengths of two or more architectures to produce a hybrid that better encompasses human phenomena. A simple example, the combination of Soar's cognitive processor with EPIC's perceptual and motor processors, has already been mentioned. In contrast with ACT-R/PM's reimplementation of EPIC-like processors in the ACT architecture itself, neither Soar nor EPIC was rewritten, but communicate through a shared working memory. More fundamental combinations of architectures are the subject of an ongoing basic research program at the Office of Naval Research (Hybrid Architectures as Models of Human Learning), which supported the infancy of several hybrid architectures. To address the effects of environmental frequency in Soar, that architecture was combined with Echo (Thagard, 1989), a statistical technique for belief updating. Neural nets were augmented with a symbolic explanation-based learning system (Dietterich and Flann, 1997) to address the learning of long procedures when the effects of actions are widely separated from the actions, credit and blame are difficult to assign, and the right combination of moves should count more toward learning than the myriad of failures along the way. The CLARION architecture (Sun, 1995) integrates reactive routines, generic rules, learning, and decision making to develop versatile agents that learn in situated contexts and generalize resulting knowledge to different environments. Gordan (1995) extends Marshall's schema theory of human learning, supplementing its high-level planner with a low-level stimulus-response capability. Marshall (1995) is using a neural net for the identification component of schema theory and ACT-R for the elaboration, planning, and execution components. Cohen and Thompson (1995) use a localist/connectionist model to support rapid recognitional domain reasoning, a distributed