"start-up" time, after which only much slower tracking will be sufficient. This can be achieved by using a) smaller size blocks at the start and then gradually increasing the block length ancl/or b) smaller soft-constrained parameter at the beginning and then increasing this parameter. The latter algorithm is referred to as "growing memory tracking" in [3]. However, due to the modifications to the block-LS cost function, the optimality is traded in favor of information transfer between the data blocks.We will begin by showing how the soft-constrained term in [3] can be chosen to guarantee a near optimal solution. Then, a new modified block FTF algorithm is suggested that provides a more efficient means of weight transferring between the data blocks while maintaining the optimality of the solution. The implementation of this scheme using the original block FTF structure is then suggested. The effectiveness of our algorithm is tested on the underwater target identification problem from acoustic backscattered signals.Various signal processing schemes [4]- [8] have been proposed in the literature to extract resonance information from the acoustic backscatter from objects of regular shapes such as thin spherical or cylindrical shells submerged in water. In [4], a joint time-frequency analysis of the impulse response of a spherical shell has been studied for characterization of the surface waves on an elastic shell using the Wigner-Ville distribution. The wavelet transform using five-cycle cosinemodulated Gaussian wavelet approximations was applied to the impulse response of a spherical shell of differing thickness to examine resonance characteristics of the target [5]. A wavelet-based classifier that uses an artificial neural network to adaptively compute discriminatory information on a target in the form of locations, sizes, and weights of Gaussian patches in time scale is described in [6]. The method in [7] uses a shorttime Fourier transform to determine the resonance spectrum of submerged elastic cylindrical wires.In [8], an RLS-based adaptive filtering scheme for the separation of resonant and specular components in the backseattered signal from objects of arbitrary shape is developed. The test results in this reference indicated that unlike the previous methods, no underlying model assumptions on the elastic return, and no a priori knowledge on the signal statistics would be required. In addition, this method offers excellent robustness in presence of noise and great performance for relatively stationary backscattered signals and is ideally suited for on-line processing situations. However, these results also revealed the interesting fact that for certain aspect angles, where the backscattered signal exhibits highly nonstationary 1053-587X!96$05.00