Robust Solutions
Robustness is best understood not as a property of a problem solver,
but of the solutions that it produces. It is not that a
planner is robust, but that a planner produces
robust plans. From a practical point of view, the fact that a
plan is capable of surviving a variety of situational changes or
adversary's responses (i.e., is robust) is far more important than the
fact that the planner is capable of adapting plans to situational change:
the luxury of time to replan may not be available.
CIRL
We have identified related properties of solutions that lead to robust
plans. In particular, plans that correspond to a
solution cluster are robust in
defined and determinable ways. Similarly, by specifying that
solutions must be supermodels,
it is possible to construct solutions that have specified levels
of robustness (if such solutions exist for the problem at hand).
We have discovered that solution clusters and supermodels exist
for a variety of problems, and are now exploring mechanisms for
finding and exploiting them. Important problems include finding
ways of specifying limited robustness (i.e., robust against threats
of certain types, not all types), and specifying the range of
acceptable ``fixes'' under which a solution can be considered robust
(e.g., you can't change what you did yesterday).
Pointers
- Parent areas:
- Incomplete Knowledge
- Model Clustering
- Supermodels
- Implementation areas:
- SWFM
- O-Plan
Back to the CIRL Overview Page.