Accuracy and Generality of Trained Models for Lift Planning Using Deep Reinforcement Learning-Optimization of the Crane Hook Movement between Two Points

An optimization system of a lifting plan must have generality to manage different design conditions and real-time changes at a construction site. Furthermore, it must optimize the construction planning and scheduling. In a previous study, we trained a two-point locomotion model for crane hook movement using deep reinforcement learning to generate and optimize lifting plans automatically. However, we did not test the accuracy and generality of the model. In this study, we test (1) the accuracy and (2) the generality of the trained model using a new environment. To evaluate the accuracy of the optimal solution, we examined the locus of the movement of each frame between two points. To verify the generality of the trained model, we solved an optimization problem of the crane hook movement under different conditions of the crane's learning environment using the trained model. From the results, we found that the movement path was 3.6 times the shortest path and the crane hook initially moved vertical. Furthermore, the agent solved the optimization problem of the crane hook movement when the size of the crane changed. Therefore, the corresponding range increased with increasing size of the crane. However, the agent did not solve the problem when the slewing angle in the target position was larger than the slewing angle in training. Based on these results, we believe that the limited vertical movement range and rotation range of the crane reduces the accuracy and generality of the trained model.