As a kind of electromagnetic exploration method in frequency domain, spectral induced polarization (SIP) technique is a branch of electromagnetic exploration methods in geophysics exploration, which is widely used in environmental and engineering geophysical prospecting, as well as mineral exploration, oil & gas and coal exploration. In the theory of SIP data interpretation, inversion of three spectral induced polarization parameters (m, τ and c) in Cole-Cole model is a basic job. Due to their respective special character in the function expression, the inversion of m, τ and c from frequency spectral response data based on linear inversion theory is usually instable and even sometimes not convergent when data have a little errors. In this paper, we adopt a kind of improved genetic algorithm to implement inversion. Inversion on one Cole-Cole model shows that the algorithm converges fast, has very good stability and high precision, and even can permits the observed data random error up to 15%. Inversion on two Cole-Cole models, which simulates the fact that the effects of electromagnetic and induced polarization exist simultaneously, show the algorithm is still fast convergence and high precision, and permits a few observed data error (say, 5% random error). Only when random errors add up to 10%, the results of inversion begin to become obvious error.