information when it becomes
available. The aforementioned sequential decision methods cannot readily cope
with action and state sets that are only partially known or expanding or shrinking
over time, nor with only partially known or evolving transition probabilities. Off
the shelf, dynamic programming does work, unabridged, with a transient, nonergodic
transition matrix and hence (a narrow form of) dynamics, yet it does
require the state space and transition matrix to be constant.