As the demand for high-performance electronic devices has increased, the semiconductor manufacturing process is being developed centering on the production of multi-chip products. In multi-chip products, re-entrance occurs by repeating the process several times in the packaging line, and the setup change of equipment is frequently incurred. These are major factors that make the scheduling of the semiconductor packaging line difficult. The production environment frequently changes due to internal and external variabilities. In addition, since the calculation time required for scheduling is very important at the manufacturing site, prompt schedule generation is required. As the research of the semiconductor packaging line scheduling becomes active, the reinforcement learning based scheduling research aiming at the global optimization is increasing. In view of the utilization of scheduling research based on reinforcement learning, there is a need for a method capable of reacting to various production environment changes and obtaining a good schedule in a short time.This study aims at obtaining the robustness of the scheduling model based on deep reinforcement learning. We propose a regularzied training method for semiconductor packaging lines scheduling based on deep reinforcement learning without performance degradation and re-training when a new production environment is given as a test data. In order to apply reinforcement learning to flexible job-shop scheduling problem, we designed state, action and reward considering overall process and trained deep Q network which is a representative algorithm of deep reinforcement learning. The regularzied training method proposed in this study is divided into four stages and designed to train the generalities of the problems reflected in various production environment and the specificity of each problem. Experiments were conducted using scheduling problems of different complexity, and it was verified that the performance was superior to other scheduling models based on rule-based and deep reinforcement learning.This study is the first research that focuses on the robustness of the model in the reinforcement learning based scheduling. Moreover, the result of this study enhances the practicality of research in real factory application.