In this paper. we introduce a method of designing optimal time-frequency de
tectors from training samples, which is potentially of great benefit when f
ew a priori information on the nonstationary signal to be detected is avail
able. However, achieving good performance with data-driven detectors requir
es matching their complexity to the available amount of training samples: r
eceivers with a too large number of adjustable parameters often exhibit a p
oor generalization performance whereas those with an insufficient complexit
y cannot learn all the information available in the design set. Then, using
the principle of structural risk minimization proposed by Vapnik, we intro
duce procedures which provide powerful tools for tuning the complexity of g
eneralized linear detectors and improving their performance. Next, these me
thods are successfully experimented on simulated and real data, with linear
detectors operating in the time-frequency domain: it is in such high-dimen
sional feature spaces thar procedures of deriving reduced-bias receivers fr
om training samples are of prime necessity. Finally, we show that our metho
dology may offer a helpful support for designing detectors in many applicat
ions of current interest, such as biomedical engineering and complex system
s monitoring. (C) 1999 Elsevier Science B.V. All rights reserved.