GBDT Modeling of Deep Reinforcement Learning Agents Using Distillation

In recent years, the field of deep reinforcement learning has been developed, and it has become possible to learn complex task processing automatically. However, it is difficult to verify the safety of deep reinforcement learning methods because their contents are not clear, and it is difficult to find a way to improve them. Therefore, in this study, we aimed to transform the solution obtained internally by deep reinforcement learning into a GBDT model using distillation, which is commonly used in the imaging field. This method enables us to confirm the validity of the obtained model by converting the solution represented by the neural network into a GBDT model, which reduces the number of calculations and speeds up the processing time.