Application of q-learning algorithm to bicriteria dynamic scheduling problem

Most of research in manufacturing scheduling is concerned with the minimization of a single criterion. However, scheduling problems often involve more than one objective and therefore require multi criteria analysis. This thesis deals with bicriteria dynamic scheduling problems. The main purpose of this study is to find out a compromising solution for the system objectives. The Q-learning algorithm, an agent based approach, is proposed to find a good schedule for the systems. The proposed methodology consists of three phases. In the first phase, selected dispatching rules are performed on the system under different conditions for both of the objectives. In the second phase, learning potential of the Q-agent is investigated. In the third phase, Q-learning agent is trained to minimize the system objectives on dynamic scheduling problem. Dispatching rules which are involved in the learning process are determined according to results in the first phase. In this thesis, bicriteria dynamic scheduling problem is investigated on a single machine and a flow shop separately. Furthermore, the performance of the Q-learning agent is evaluated and compared to other dispatching rules by using a new ranking method which is also presented in this thesis. Keywords: Reinforcement learning