The classical problem of quickest change detection is studied with an
additional
constraint on the cost of observations used in the detection process. Bayesian
and Minimax formulations are proposed for the problem. It is shown that
two-threshold generalizations of the classical single-threshold tests are
asymptotically optimal for the proposed formulations. Guidelines are given for
the design of the proposed algorithms. Numerical results are provided that
reveal that the proposed algorithms have good trade-off curves,
and perform significantly better than the approach of fractional sampling,
where the constraint on the cost of observations is met by skipping
observations
randomly. Extensions of this work to sensor networks is also discussed.