Ambulance Redeployment via Reinforcement Learning

In this study, ambulance redeployment is performed by using reinforcement learning methods. The objective in the ambulance redeployment problem is to redeploy the limited number of ambulances in such a way to minimize the arrival times to calls. Good resolution of the ambulance redeployment problem is crucial to the development in a country's emergency medical services and to saving human life during emergencies. Contrary to the commonly used optimization methods in the literature, a learning-based method is used to solve this problem. Therefore, ambulance redeployment can be performed successfully even in cases where the case distributions are not known in advance and when traffic on the roads changes with respect to time. In the solution of the problem, better ambulance waiting locations are computed using the trade-off between exploration and exploitation. The proposed algorithms calculate new waiting locations for ambulances at regular intervals using the call distributions and traffic information that are observed so far. During the testing phase, the proposed algorithms are compared against the oracle optimization algorithm, which performs static allocation of the ambulances and knows the call distributions beforehand. Under the same conditions, it is shown that the proposed algorithms perform similarly to the oracle optimization algorithm.