This paper introduces directional interchangeability, a weak form of neighborhood interchangeability [6]. The basic idea is
that although two values of a variable may not be neighborhood interchangeable if we consider the whole neighborhood of the
variable, they could be neighborhood interchangeable if we restrict the neighborhood to a subset of neighboring variables
induced by a variable ordering.
In spite of the fact that the proposed concept cannot be used to remove redundant values while preserving problem satisfiability,
it provides a mean to partition value domains into subsets of directionally interchangeable values that can be attempted simultaneously
by a tree search.
Several experiments carried out on various binary CSPs, clearly indicate that variations of the Forward-Checking algorithm
and the Maintaining Arc-Consistency algorithm that exploit directional interchangeability often outperform the original algorithms.