| Abstract [eng] |
Facility location problems involve strategic decisions with long term economic and operational consequences, particularly in competitive markets where customer behavior plays a central role. In competitive facility location, optimal placement depends not only on spatial demand and competitor locations, but also on assumptions regarding how customers choose among alternatives. Small changes in these assumptions can lead to significantly different solutions, making models optimized under a single behavioral specification fragile. This thesis investigates the application of deep reinforcement learning to competitive facility location under uncertainty in customer behavior. The problem is studied from the perspective of an entering firm that sequentially places new facilities in a market with existing competitors, aiming to maximize captured market share. Customer choice is modeled using Pareto-Huff filter, while demand allocation follows either binary or proportional rules, representing two behavioral assumptions. A deep Q-learning framework is developed and trained across a diverse set of simulated instances. Robustness is evaluated empirically by assessing performance under both customer behavior types and selecting a representative policy using Manhattan distance to complete market share capture. Experimental results show that the learned policies achieve competitive market share outcomes while offering substantial computational advantages over exhaustive enumeration. Although performance varies across problem settings and does not consistently reach optimality, the results indicate that reinforcement learning provides a promising and scalable foundation for competitive facility location under behavioral uncertainty. |
