Local Failure vs. Bayesian Methods for Finding Stable Solutions to Dynamic Constraint Satisfaction Problems

In recent years the constraint satisfaction paradigm has been extended to the case in which a problem can undergo changes that may invalidate the current solution. If such changes are recurrent, it may be possible to nd solutions that are more likely to remain valid after such changes (stable solutions). In earlier work we have found that, (i) hill climbing procedures such as min-connicts are peculiarly suitable for dealing with dynamic CSPs, in part because, (ii) these procedures can be readily combined with simple strategies for penalizing values that`go bad' through simple loss or changes that render them invalid as assignments. However, the situation we have studied also lends itself to a Bayesian approach for choosing values that are more likely to yield stable solutions. In comparing the two approaches, we nd that the Bayesian methods are less eecient with respect to runtime, and, more surprisingly, that the solutions obtained are no more stable than those found with penalty procedures under comparable conditions. Relative eeciency depends on not preventing hill climbing from nding any solution because there are no suitable candidate values, and this is accomplished more eeectively by penalty methods. Diierences in solution stability depend not only on the way in which values are diierentiated, but also on the manner in which values are selected for consideration during the course of search.