Route optimization of construction machine by deep reinforcement learning

After it was reported that an AI player scored higher in Atari2600 games than skilled human players by using deep reinforcement learning techniques, many researchers were inspired to apply deep reinforcement leaning in various fields. This paper focuses on the autonomous ground leveling work by a bulldozer, which is expected to optimize the action of the bulldozer. In a previous work, we implemented a deep Q learning method by giving the images as the input data for the network. However, when learning the image using the convolution layer as the input using deep reinforcement learning, it requires a large computational cost for the learning process. If the size of the neural network is shrunken by contriving the data to be supplied to the input, the learning time (duration) will be reduced. This paper describes the comparison results for different orders of input data. the transition of the learning sequence is also evaluated.