Environments for Soar

• Knowledge-Rich Environments

  Soar is able to use knowledge from a variety of sources simultaneously due to the parallel nature of the elaboration phase. However, problem spaces also provide a natural compartmentalization of unrelated knowledge sources. The combination of these two considerations makes Soar ideal for problems in which a lot a knowledge is available, avoiding overloads when there is a preponderance of data and knowledge.

• Complex Environments

  The compartmentalization of knowledge provided by problem spaces also gives Soar the ability to manage the complexity of the environment even though it has limited resources. Some additional ability may come from match-based attention and the ability to recognize familiar events in the midst of complex inputs via the associational character of long-term memory.

• Search

  Soar was designed as a mechanism to realize the weak methods. As such, it is ideal for problems that involve search. Problem Spaces provide a definition of the search space and the operators that may be applied. Chunking allows knowledge gained during search to be cached, resulting in problem solving that is initially search intensive but that becomes more routine as problem solving progresses.

• Learning

  Soar's universal learning mechanism, chunking, makes it particularly applicable to problems in which a considerable performance improvement can be realized through learning search control or general problem knowledge. Learning cost is minimized in Soar because the impasse mechanism provides the necessary knowledge about when and what to learn.

• Dynamic Environments

  Encoding and decoding productions allow Soar to monitor an immense amount of changing detail. The match-based attention mechanism of the production system allows salient behavior even in the midst of this perceptual diversity. Thus, Soar is able to maintain its reactivity in dynamic environments. One limitation here is that the environment should require reaction no faster than the decision phase.

• Static Environments

  Soar can act as a general problem solver and is just as applicable to environments which are stable as to dynamic ones. In particular, Soar has been applied to many games (such as the 8-puzzle and Towers of Hanoi) which have very static environments (nothing changes except through the direct action of the agent). Although Soar may not be as efficient as special-purpose techniques for the particular problem, it may certainly be applied and achieve success in these domains.

• Unpredictable and Uncertain Environments

  Soar is chiefly able to operate in uncertain environments due to its utilization of a minimum commitment strategy. This allows the specifics surrounding an event to be instantiated at run-time rather than being necessarily anticipated beforehand. Additionally, the generalization of chunking also alleviates the need to be certain of forthcoming events; if the events are similar in scope (or if a variety of different events have been experienced by the system), the learned chunks will apply to the new situation.

• Soar Requires Consistent Environments

  Unfortunately, Soar requires consistent environments, ones whose properties do not change over time. This is a direct result of the impenetrability and permanence of productions. An additional problem is that learned chunks also reside in this memory, resulting in improper (and perhaps erroneous) behavior when the appropriate response to the same environment changes significantly. This problem is currently being addressed and is discussed among the issues.