One of the tacit aims of any field in science is the unification of
its theories and laws. For example, the field of physics has this
far-off aim of finding unifying equations to explain all physical
phenomena. Allen Newell, in his book, Unified Theories of
Cognition, urges the AI and cognitive science communities to
endeavor to develop unifying theories for cognition. As in physics,
unified theories of cognition will serve to unify all existing
understanding of cognition. Specifically, they will be theories that
integrate a set of mechanisms from which all cognitive behavior can
emerge. Cognitive behaviors of concern are:
- problem solving, decision making, and routine action
- memory, learning, and skill
- perception and motor behavior
- language
- motivation and emotion
- imagining, dreaming, daydreaming, etc.
Newell has proposed Soar as a candidate unified theory of cognition.
Soar is a collection of mutually exclusive mechanisms that combine to
produce a system that has been shown to be applicable to a wide array
of AI (eg: planning, control, learning) and cognitive modeling
(eg:power law, reaction times) tasks.
The development of Soar has been driven by four methodological assumptions:
- there is high utility in focusing on the cognitive band as opposed to the neural or rational
bands because any complete model of general intelligence must have as
its foundation a solid model of the cognitive band on which higher
level can be built..
- general intelligence can most usefully be studied by not making a
distinction between human intelligence and artificial intelligence.
This allows a wider range of research methodologies and data to be
used to mutually constrain the structure of the system.
- as stated above, the architecture should consist of a small set of
orthogonal mechanism. All intelligent behavior should emerge from
these mechanisms. This assumption biases Soar to a simple, uniform
design rather than a toolkit approach.
- the collection of mechanisms should be pushed to their absolute
limit. In particular, one shouldn't automatically introduce a new
mechanism to achieve a desired capability until it has been
empirically shown that the existing mechanisms cannot be, shall we
say, coerced into producing the desired capability.
Characteristics
Sources and References
Laird, J.E., Newell, A., and Rosenbloom, P.S., "SOAR: An Architecture
for General Intelligence" in Artificial Intelligence, vol. 33
(1987), pp. 1-64.
Laird, J.E., and Rosenbloom, P.S.,"Integrating Execution, Planning,
and Learning in Soar for External Environments", in Proceedings of the
Eight National Conference on Artificial Intelligence (AAAI-90),
August, 1990.
Laird, J.E., Rosenbloom, P.S., and Newell, A., "Chunking in Soar: The
Anatomy of a General Learning Mechanism" in Machine Learning,
vol. 1 (1986) pp. 11-46.
Laird, J.E., Yager, E.S., Hucka, M., and Tuck, C.M., "Robo-Soar: An
Integration of External Interaction, Planning, and Learning using
Soar", Robotics and Autonomous Systems, vol. 8, pp. 113-29.
Laird, J.E., Hucka, M., Huffman, S., and Rosenbloom, P.S., "An
Analysis of Soar as an Integrated Architecture," SIGART Bulletin
2,1991, pp. 98-103.
Laird, J.E., Congdon, C.B., Altmann, E., and Doorenbos, R., "The Soar
User's Manual, Version 6, Edition 1."
Newell, A., "Unified Theories of Cognition", Harvard Press, 1990.
Rosenbloom, P.S., Laird, J.E., Newell, A., and McCarl, R., "A
preliminary analysis of the Soar architecture as a basis for general
intelligence" in Artificial Intelligence, vol. 37 (1991),
pp. 289-325.
Rosenbloom, P.S., Newell, A., and Laird, J.E., "Toward the Knowledge
Level in Soar: The Role of the Architecture in the Use of Knowledge"
in Kurt VanLehn (Ed.), Architectures For Intelligence, pp.
75-111, Lawrence Erlbaum Associates, Hillsdale,NJ, 1991.
To learn more about Soar and the Soar community, please take a peek at their
WWW page.