Imperfect Knowledge
If the knowledge the agent has is not completely accurate and
complete, it is said to be imperfect. There are many ways in which an
agent can acquire imperfect knowledge. In a
dynamic environment, something may change without the agent's
control. This may invalidate some of the agent's knowledge. The
agent may or may not be aware of the change or the invalidation. The
real world is very unpredictable. Sensors
and actuators may not be perfect. This can lead to incorrect
perceptions and uncertainty about whether actions are performed.
There is a wide range of ways to deal with imperfect knowledge. One
extreme is not to maintain any world knowledge. This prevents one way
of acquiring imperfect knowledge from even occuring. The architecture
can also ignore the possibilty of imperfect knowledge due to sensors and actuators . If it maintains world
knowledge, consistency
becomes an issue. Another method is assign probabilities to every
piece of knowledge. It could also assume that you will have imperfect
knowledge and try to correct for it when it affects the system
(failure awareness).
Architectures which do not maintain world knowledge:
Architectures which do maintain world knowledge:
Architectures which ignore sensor and actuator error:
Architectures which try to compensate with probabilities:
Architectures which try to compensate with failure awareness:
- Icarus (P. Langley and others)
- ATLANTIS (E. Gat)
- SOAR (J. Laird, P. Rosenbloom, and others)
- ERE (M. Drummond and others)
- Guardian (B. Hayes-Roth)
Architectures which try to compensate with a truth maintence system:
- Theo (T. M. Mitchell and others)
- TETON (K. Vanlehn and W. Ball)
- SOAR (J. Laird, P. Rosenbloom, and others)
- HOMER (S. Vere and T. Bickmore)
Architectures which try to compensate with other methods:
- Theo (T. M. Mitchell and others)
Other characteristics of the Environment and
Agent Body
Back to the Title Page