A simple change detection scheme

International audienceThe problem of detecting changes in a stochastic system is addressed. When the model parameters after the change are unknown the generalized likelihood ratio (GLR) scheme is usually used to solve the problem. This scheme is asymptotically optimal but it is also particularly time-consuming which makes questionable its real time implementation. The window-limited GLR scheme, which takes into account only significant (for the detection) previous observations, is less demanding but often it is still time-consuming. In this paper we introduce an alternative approach to reduce the computational burden of the GLR scheme. The idea of this solution is to decompose a given parameter space into several subsets so chosen that in each subset the detection problem can be solved with loss of a small part of optimality by a recursive change detection algorithm