Reinforcement learning for urban business intelligence

Official URL: https://dx.doi.org/10.1109/TCSS.2026.3692595

The importance of location for business success cannot be overstated. Existing approaches to the business location selection problem often involve creating extensively tuned models specific to the geographical and economic climate being analyzed, making them difficult to adapt to other scenarios. This poses a significant challenge in regions where data are scarce or costly to obtain. In this article, we experiment with various offline deep Q-learning (DQL) frameworks for business location recommendation that can be trained in one geographic area and applied to another without requiring further training or tuning. Our comprehensive experiments on real-world data demonstrate the superior generalizability of our DQL approaches, outperforming the well-established Huff gravity model by more than 25% in profit realization on average across all DQN variants and business categories, with an average profit realization of 75.35%. Our results indicate that the variation in the training data must be as high as the variation in the test data for the model to be successfully applied to other locations, despite discrepancies between the characteristics of the cities. Our approach offers a highly generalizable and easily applicable solution to the business location selection problem, providing a strong alternative to gravity-based models in settings with limited access to localized data.