The multi-collinearity analysis and the extreme gradient lifting model were used to determine the impact factors of mountain torrent disaster vulnerability of Hainan Island, and a reinforcement learning (RL) evaluation model of mountain torrent disasters was proposed based on the SAC (soft actor-critic) RL algorithm. The RL model results were compared with those of the convolution neural network (CNN) and deep neural network(DNN) models based on the sensitivity and specificity, as well as the receiver operating characteristic curve. Results show that the RL model has higher accuracy and reliability, and its area under curve (AUC) index of the test data set reaches 0.931, which is higher than those of the CNN and DNN models. The RL model can provide scientific basis for mountain torrent disaster risk assessment in Hainan Island.