models are natural-language comprehension (Lewis, 1996, 1997a, 1997b), syllogistic reasoning (Polk and Newell, 1995), concept acquisition (Miller and Laird, 1996), use of a help system (Peck and John, 1992; Ritter and Larkin, 1994), learning and use of episodic information (Altmann and John, forthcoming), and various human-computer interaction tasks (Howes and Young, 1996, 1997; Rieman et al., 1996). In addition, other work that compares Soar models with human behavior include models of covert visual search (Weismeyer, 1992), abductive reasoning (Krems and Johnson, 1995), a simplified air traffic controller task (Bass et al., 1995), the job of the NASA test director (Nelson et al., 1994a, 1994b), job-shop scheduling (Nerb et al., 1993), and driving a car (Aasman, 1995).
Recent work has joined the Soar cognitive architecture to EPIC's perceptual and motor processors (Chong and Laird, 1997). The resulting perceptual-cognitivemotor model has been validated against human dual-task data and shown to be as accurate as EPIC alone, despite the differences between the cognitive processing of EPIC and SOAR. This approach of incorporating the mechanisms of other architectures and models and "inheriting" their validation against human data promises to result in rapid progress as parallel developments by other architectures emerge.
Soar has been used extensively for military problems. Most notably, the agents of Soar-intelligent forces (Soar-IFOR) have been participants in simulated theaters of war (STOW-E in 1995, STOW-97 in October 1997). These agents have simulated the behavior of fixed-wing and rotary-wing pilots on combat and reconnaissance missions, competing favorably with human pilots. They have acted as very large expert systems, using Soar's architecture to encode knowledge prior to performance and then acting on it, or solving problems using that knowledge. They use this knowledge to act as individuals and as groups of cooperating individuals. However, these systems have not yet run with learning capabilities.
Other research Soar systems have demonstrated more far-reaching capabilities and would require some additional research before becoming robust, real-time systems for military use. An example is the system noted above that integrated natural-language capabilities, visual search, and task knowledge to model the decision process of the NASA test director in simulated real time (slower than actual real time) (Nelson et al., 1994a, 1994b). Another system used ubiquitous learning to store and recall situations and decisions during operation for post-mission debriefing (Johnson, 1994). Although some work has been done on modeling frequency learning in laboratory tasks (Miller and Laird, 1996), it is doubtful that Soar's purely symbolic learning mechanism will extend far enough to capture human-like adaptivity to the frequency of occurrences in real-world environments. In addition, as with most of the other architectures reviewed here, no actual modeling has been done to account for the effects of moderator variables,