Self-learning modeling is a technique used to automatically learn and optimize model parameters from data, which is significant for modeling in process of possibilistic model checking. Self-learning modeling based on generalized possibility Kripke structure (GPKS) has been studied to some extent. Recently, model checking based on generalized possibilistic decision process (GPDP) has become a research focus. However, the lack of effective self-learning methods for constructing GPDP models limits the practical applications of GPDP-based model checking. For complex GPDP models, manually generating fuzzy states and possibility transitions is extremely time-consuming and resource-consuming. To address this issue, we developed an online supervised learning algorithm capable of learning the fuzzy states and possibilistic transition matrices of a GPDP. By using Gaussian fuzzy functions to associate external variables with fuzzy states, our new GPDP model with fuzzifier sets (GPDP-FS) maps the values of external variables to the atomic proposition values of fuzzy states in the GPDP model. For the action set of the GPDP model, we set the transition mechanism of each action in the action set as different atomic proposition evolution matrix. The algorithm uses stochastic gradient descent to emulate the state transition process and build the possibility transition matrix which is an essential component of model checking. Results indicate that our method can derive fuzzy states, state transition mechanisms, and possibility transition matrices solely from external variables. This approach enables GPDP modeling through a self-learning process, overcoming the modeling problem of possibilistic decision processes in practical applications.