Paper Type
Complete
Paper Number
PACIS2026-1291
Description
Urban parking operators typically serve two customer groups with competing objectives: public customers (e.g., visitors paying hourly fees) generate short-term revenue, while registered customers (e.g., permit holders or long-term subscribers) expect reliable access. Allocating limited capacity between these groups under uncertain arrivals and parking durations creates a dynamic decision problem. We formulate this setting as a Markov decision process with an aggregated (partially observable) state representation and train a reinforcement learning agent in a simulation parameterized from historical operational data. The learned policies dynamically adjust reservation levels in response to evolving demand conditions. Results reveal a clear revenue-service trade-off. Rather than identifying a single optimum, the approach characterizes a Pareto frontier of allocation strategies. Compared to fixed reservation ratios and reactive buffer heuristics, learned policies expand the attainable trade-off region, achieving higher revenue for comparable service levels in practically relevant operating regimes.
Recommended Citation
Müller, Thomas; Schweim, Dirk; and Rothlauf, Franz, "Balancing Revenue and Service in Urban Parking: A Reinforcement Learning-Based Capacity Management Approach" (2026). PACIS 2026 Proceedings. 2.
https://aisel.aisnet.org/pacis2026/iot_smartcity/iot_smartcity/2
Balancing Revenue and Service in Urban Parking: A Reinforcement Learning-Based Capacity Management Approach
Urban parking operators typically serve two customer groups with competing objectives: public customers (e.g., visitors paying hourly fees) generate short-term revenue, while registered customers (e.g., permit holders or long-term subscribers) expect reliable access. Allocating limited capacity between these groups under uncertain arrivals and parking durations creates a dynamic decision problem. We formulate this setting as a Markov decision process with an aggregated (partially observable) state representation and train a reinforcement learning agent in a simulation parameterized from historical operational data. The learned policies dynamically adjust reservation levels in response to evolving demand conditions. Results reveal a clear revenue-service trade-off. Rather than identifying a single optimum, the approach characterizes a Pareto frontier of allocation strategies. Compared to fixed reservation ratios and reactive buffer heuristics, learned policies expand the attainable trade-off region, achieving higher revenue for comparable service levels in practically relevant operating regimes.
Comments
10-IoT