I am interested in composite hypothesis testing where one can not easily estimate the nuisance parameter of a particular hypothesis even when the data size becomes large. A typical situation occurs in detecting weak sinusoidal type of signal buried by heavy noise floor (SNR<-20dB) with random phase reversal due to the flip of spin alignment in magnetic resonance force microscopy. Another application is to detect weak echoes from target return through the processing of multiple radar scans. Traditional track-before-detect method relies on the assumption that the target, if exists, undergoes a certain deterministic motion which can be extracted out through some kind of “smart” measurement alignment with a proper account for the possible clutter originated measurements. However, a target can be evasive which makes its motion highly unpredictable and the generalized likelihood ratio test won’t work in this case simply due to the unreliable estimate of target motion parameters.

 

Traditionally, hypothesis testing is treated within a parametric setting of the data generation assumption, e.g., the likelihood function of each hypothesis can be fully specified given the possibly unknown nuisance parameter. As the data size grows and the advances of statistical learning theory become more and more accessible to the engineers, the need to handle signal detection problem in a nonparametric or even semi-parametric setting draws much of the attention to me (and I believe to lots of researchers in this field). I have been working on model selection problems for determining the order of linear regression models when the noise distribution is unknown. I see a strong connection to the inherent difficulty of making a composite hypothesis simple when little can be said about the nuisance parameter. Besides, even the performance evaluation of algorithms operating on multiple composite hypothesis testing is a nontrivial issue. I would like to build a semi-parametric framework to treat those nuisance parameters separately, i.e., either estimable or non-estimable, for weak signal detection problems where the generalized likelihood test fails to yield good performance.

 

Sadly, I have not made much progress so far but I already spotted the non-optimal behaviors of some popularly used approaches in hypothesis testing and model selection. I hope that people will gradually switch their focus to the formulation of the hypothesis testing itself from trying to develop a so called optimal test procedure under certain restrictive assumptions. My research has been supported in part by DARPA MOSAIC, NSF/LEQSF-PFund.

--------------------------------------------------------------------------------------------------------------------------------------------

People working in Signal Detection area

 

H V Poor at Princeton, S A Kassam at UPenn, S M Kay at URI, P K Willett at UConn, Harry Van Trees at GMU, L Scharf at UColorado