Soar Issues Deriving from Knowledge, Memory and Representation

• Efficiency of the Matcher

  The production match during the elaboration phase requires the greatest amount of computational resources for Soar systems. Although this process is implicitly parallel, it is normally instantiated on serial computers. As the size of the knowledge base grows, Soar systems may exhibit a concomitant slow-down related to inefficiencies in the match algorithm. Additionally, although the RETE algorithm used in the Soar production system is quite efficient, there are physical limits on the speed of the production match. Thus, the scalability of Soar systems is an open question for systems with very large knowledge bases.

• Taskability

  Although the taskability of Soar agents has improved with the addition of a natural language capability, Soar systems are not very taskable at run-time. This problem derives, in part, from Soar's explicit use of a goal stack. Giving Soar a new task requires manipulation of this goal stack. Recent work (see, for example, TacAir-Soar) has demonstrated that taskability may be achieved naturally within a framework that supports (and expects) outside direction.

• Goal Reconstruction

  Soar currently provides no architectural mechanism for goal reconstruction. In particular, all subgoals remain on the context stack as long as the impasse which caused the creation of the subgoal is unresolved. Although this may seem a small problem, especially for short-term goals, for longer time periods goal reconstruction seems more natural than keeping a goal on the stack. For example, Soar would keep the goal of "going to California next year" explicitly on its goal stack from the time the goal was first considered until it was resolved. Such a method has repercussions for both long-term, continuous use of Soar systems and Soar's psychological validity.

• Multiple, Simultaneous Goals

  A related problem to the one discussed above is that the Soar does not support the consideration of interacting and conflicting goals. All goals are ordered on the goal stack in an explicit hierarchy. Since Soar may bring knowledge to bear for any slot in this context stack, conflicting information may be considered simultaneously. Although this is neither a problem for the monotonic production system or the impasse-driven decision procedure (i.e., an impasse will simply arise when the conflicting information is considered), the interaction of these goals will continue to generate impasses when a simple solution may exist that represents a compromise of the two competing goals.

To return, press HOME.