Non-routine manual operations of a plant require appropriate judgment according to the operating state, and significantly depend on the knowledge and skill of the operator. Deep reinforcement learning can recognize a state and learn a series of operations based on an evaluation index for that state. It is expected as a method that leads to the clarification of operational guidelines, which have been considered tacit knowledge thus far. However, due to the problems such as the difficulty in designing the evaluation indexes required for learning, the explainability of the learning results, etc., its application to plant operations, which requires safety and reliability, has not progressed. This report presents the acquisition of expert operations using inverse reinforcement learning technology developed in collaboration with Chiba University and the technology to surmise the know-how of experts through visualization of the learning results.
