As a major production facility for oil and gas exploration, a dangerous working environment can cause serious accidents in FPSOs. The hull of the FPSO can be severely damaged or flooded, and in such cases, it is very difficult for workers to escape due to the large and complex structure of the FPSO. Even in an emergency situation, In an emergency situation, rational reasoning is blurred, and objective and prompt judgment cannot be made. It is necessary to provide quick and efficient evacuation guidance in case of emergency in order to prevent casualties. In this study, deep Q-network, a method of reinforcement learning, was applied to the optimal path prediction model to calculate the optimal evacuation route for workers in case of an accident. Deep Q-network can be applied to models with large and complex structure, and the environment of the model consists of 4 parts: deck, accommodation, backward, and frontward. The agent receives penalties for every move and is rewarded when it arrives at one of the four exits on the deck. As a result, the average number of movements is less than 30 when escaping from all locations in the environment consisting of 621 grids. This study contributes to provide an optimal escape route that is fast and safe in a changing environment. Through the analysis of the results of this study, it is expected that the safety inspection on the FPSO can be carried out and it will help the safe FPSO design.