Investigating the multi-objective optimization of quality and efficiency using deep reinforcement learning

This study proposes a pipeline, which is based on deep reinforcement learning, aims to solve the multi-objective problem (MOP) on efficiency and quality in manufacturing. The rapid development in the area of artificial intelligent has caused a series of reactions that stirred the traditional manufacturing, pushing for the better machining quality and higher productivity. Despite all this, there has been very little research applying reinforcement learning to solve practical problems in milling process. The proposed pipeline is a two-step algorithm and makes full use of double deep Q network (DDQN) to settle the MOP of milling parameters. Firstly, surface roughness (Ra) and material removal rate (MRR) are selected as quality and efficiency indicators, respectively. In specific, the reliable prediction model of Ra is constructed on a small batch raw data via DDQN improved support vector regression (DDQN-SVR) rather than sophisticated and complex physical modeling. The MRR model is constructed by an accepted empirical formula. Then, DDQN is employed again to solve the MOP of satisfying minimum Ra and maximum MRR and compared to other accepted algorithms. Eventually, the optimal combination of machining parameters determined by entropy method was validated by experiment.