Abstarct: One of the most important data used for oil and gas exploration is reflection seismic data. Seismic inversion is one of the main steps in hydrocarbon exploration and reservoir modeling. There are various approaches for seismic inversion where series of stochastic inversion methods are the latest for this purpose. These methods work with geostatistical simulations algorithms. In this thesis, firstly, an improved direct sequential simulation (DSS) algorithm was presented in which local spatial parameters are entered in the simulation algorithm baxsed on some zonation criteria. The required computer codes of this new algorithm was then developed in MATLAB programming environment and its capability tested by using synthetic and real datasets. Secondly, new stationary and non-stationary 3D stochastic seismic inversion algorithms, baxsed on the conventional and improved DSS methods, were presented and their performance was evaluated by using both synthetic and real datasets. Finally, the variogram parameters were optimized during 3D stochastic non-stationary inversion by using the particle swarm optimization (PSO) method. The obtained results indicate that a more suitable spatial distribution of acoustic impedance values was achieved using the new improved DSS compared to those of the conventional DSS method. The results of non-stationary stochastic seismic inversions of noise free and noise contaminated synthetic datasets show a reduction of respectively 6.1 and 9.7% value of RMS error with respect to those produced by stationary stochastic inversion. Furthermore, validation of non-stationary stochastic seismic inversion results for noise-free dataset using the real data of two blind wells illustrates a decrease of the RMS error of 4.6 % in blind well 1 and 8.1 % in blind well 2 compared to those acquired by stationary inversion. Whereas for noisy dataset, there were RMS error decreasing of 12 % and 11.3 % using blind well 1 and 2 respectively. The results obtained for real dataset indicates the non-stationary inversion caused a reducing of 2.8 % in RMS error between inverted and real seismic data compared to the stationary inversion method. In addition, a comparison of the real data of a blind well in this case with the results of inversion showed the RMS error was decreased about 5.4 % compared to the stationary inversion. Moreover, inversion of noisy and noise free synthetic and real datasets by the optimized non-stationary stochastic seismic inversion method using the PSO approach showed a slight (about 1%) reduction of inversion error relative to the ordinary non-stationary seismic inversion.