Personalized treatment regimes tailored to account for individual characteristics have revolutionized the healthcare industry, offering substantial potential to maximize treatment benefits and enhance patient survival. However, the current methods for estimating individualized treatment rules in multi-treatment settings have limitations due to the lack of theoretical foundations and challenges related to model misspecification. To address these issues, we propose a novel generative learning approach called {CG-Learning}, that utilizes Wasserstein GAN to estimate the optimal decision rule that maximizes the value function for multi-treatment regimes. We derive important theoretical results of the proposed estimator including the nonasympototic error bound for the estimated optimal value and an upper bound for the probability that the estimated decision rule is not the optimal treatment option. To evaluate performance of {CG-Learning}, we conduct extensive simulation studies under various scenarios, with comparisons to existing approaches. Furthermore, we apply the proposed method to a dataset from the AIDS Clinical Trials Group for illustration.