A soft constraint allows a little 'wriggle room' rather than a simple yes/no constraint. For example, rather than saying X <=10, we might allow values of X that are a little greater than 10, but penalise this in some form in a fitness function. Soft constraints allow continuity in learning, but may be gradually hardened as learning progresses. Soft constraints can be used in techniques such as genetic algorithms or simulated annealing to combine constraint solving and optimisation.
Used in Chap. 9: page 127