It may be that values should change on multiple time scales.
My Asa H architecture is topped by a value module (see Trans. Kansas Acad. Sci., vol. 109,
no. 3/4, page 159, 2006, figure 1). This can consist of a value network like that in my 21 Sept. 2010 blog. Slow (long term) value change can be made by turning this Bayesian network into a (time varying) dynamic Bayesian network.
A module for (an additional) more rapid value change might be built from a subroutine that measures recent success (utility and utility change), a subroutine that measures current risk tolerance, a subroutine that measures recent degree of change in the knowledgebase, a subroutine that measures current system damage, a subroutine that measures current needs (like energy, hardware, etc.), etc., and a routine that selects (and resolves conflicts between) these subroutines.