This paper proposes a new identification method for the discrete-time impulse response model of a linear system from sampled input-output data. Our attention is especially focused on identification of the impulse response, which includes high-frequency components locally. The continuous-time impulse response of the system is approximated by a multiresolution neural network composed of the scaling and wavelet functions. Hence the system under study can be viewed as the weighted sum of a group of subsystems in which the scaling functions and wavelet functions are interpreted as their impulse responses respectively. Then the genetic algorithm and the AIC are introduced to select significant subsystems at each resolution level such that some redundant subsystems that are sensitive to the noise effects are discarded. It is shown through a simulation that the proposed method yields accurate estimate of the impulse response, even in the ill-condiitoned cases.