The relation between temporal-difference training methods and Markov m
odels, which was first noticed by Sutton, is explored. This relation i
s derived from a new perspective, and in this way the particular assoc
iation between conventional temporal-difference methods and first-orde
r Markov models is explained. We then derive a generalization of tempo
ral-difference methods that is suitable for Markov models of higher or
der. Finally, several issues related to the performance of mismatched
temporal-difference methods (i.e., the performance when the temporal-d
ifference method is not specifically designed to match the order of th
e Markov model) are investigated numerically.