Deep Q-Learning approaches for stochastic dynamic optimization problems

In this thesis, two different high-dimensional optimization problems, dynamic pricing and inventory control of perishable products and admission control and scheduling problems, are solved using deep reinforcement learning algorithms. Both problems require dynamic decision making in a stochastic environment. The increasing shelf life of the product in the first problem and the increasing number of heterogeneous job classes in the second problem increase the problem size exponentially. Solving a high-dimensional problem in a stochastic environment requires approximation methods. The Deep Q Learning (DQL) algorithm is used in this study to solve these problems. A solution approach, called pDQL, is proposed with problem-specific modifications. For the first problem, pDQL produces robust solutions in the face of a stochastic environment. For the second problem, it helps to avoid the positive bias due to the structure of the DQL algorithm. As a result, the proposed approach is able to find better results in less time compared to the DQL algorithm. pDQL is also compared to the Dynamic Programming (DP) algorithm and various heuristic approaches. In the first problem, the results found by the pDQL algorithm are on average 96.7% closer to the DP results, and 8.9% better than the DQL algorithm. For the second problem, the pDQL algorithm is on average 94.8% closer to the DP results and on average 2.9% better than the DQL algorithm.