Algorithm Based on Deep Reinforcement Learning for Irregular Shape Nesting Problem

Abstract

An automatic nesting algorithm based on deep reinforcement learning for shipbuilding is proposed. Automatic nesting method is desired because of the large number of parts to be placed. The nesting problem in shipbuilding is very difficult because of the relatively small area where they can be placed and the variety of shapes that can rotate freely. Since it is very difficult to manually create effective placement rules, this paper proposes an algorithm to generate rules autonomously based on reinforcement learning. To apply reinforcement learning to the nesting problem, we organized the problem as a Markov chain process. Based on deep q network, a type of reinforcement learning, we used the components of a real ship's block to learn the network parameters. The parts are represented in pixel format. It is confirmed that the present method was superior to the conventional method, and the results were comparable to those of a skilled person. However, for unlearned members, the results are inferior to those of the conventional method. This problem can be solved by relearning including unlearned components.