In order to measure vertical profiles of minor gas concentrations in the stratosphere, Improved Limb Atmosphere Spectrometers (ILAS and ILAS-II) have been developed. ILAS was the first generation sensor and made observations in 1996 and 1997. ILAS-II will measure atmospheric limb transmittance in 66 spectral bands (whereas 44 for ILAS) in the thermal infrared region by observing solar ray passed through the atmosphere. Vertical profiles of minor gases are simultaneously retrieved by a spectral fitting algorithm with an onion-peeling method for vertical profiling. This algorithm adopts a precise radiative transfer calculation and is very accurate, but usually the standard radiative transfer calculation needs huge volume of line-by-line calculations of molecular absorption to simulate theoretical limb transmittance spectra by using the HITRAN database. Methods for accelerating the algorithm have been required. In the ILAS operational program, a table look-up method, which needs an excellent computer system, was used for rapid calculations. We proposed a simplified method, which predicts the gas profiles from the measured limb transmittance spectra and vertical profiles of atmospheric pressure and temperature without iterative calculations by using a multiple regression technique. The Principal Component Expansion (PCE) is used for reducing the scale of the multiple regression model. In the training process, coefficients of the model are estimated from the previously retrieved data sets including measured limb transmittance spectra, vertical profiles of atmospheric pressure and temperature, and retrieved gas profiles. Then, the method predicts gas profiles from the newly measured limb transmittance spectra and pressure and temperature profiles. The validity of the method was confirmed by numerical simulation using the MODTRAN v.3.5 radiative transfer code. The proposed method was also applied to the actual 3474 ILAS observation data sets. The model trained by 3373 data sets well predicted the gas profiles for another 100 data sets which are selected randomly . This proposed method can be used for quick look of ILAS-II measured data and for generating the initial profiles for the operational spectral fitting algorithm.