Deep reinforcement learning in seat inventory control problem: an action generation approach

Nowadays, firms intend to use customer choice-based models instead of an independent demand paradigm to generate more revenue. In this paper, we address choice-based seat inventory control problem with stochastic demand using a deep reinforcement learning technique named Deep Q-Network (DQN). DQN can naturally address large state space problems with its integrated function approximation. However, it becomes intractable in the case of large discrete action space. To address this issue, we propose an Action Generation (AGen) algorithm. AGen is a greedy heuristic algorithm designed to be integrated into DQN to overcome the complexity of the original problem. It aims to greedily generate a set of “effective” actions to replace the original action space. This leads to the main achievement of this study which is to dramatically decrease the complexity of the solution method without negatively affecting its performance in a large-scale choice-based seat inventory allocation problem.