Cost optimal robust-schedules: Policy search in continuous action space using stochastic feedback

This paper investigates the application of stochastic approximation theory to the problem of learning optimal temporal buffers between activities in order to make a schedule more robust. We investigate this problem for the domain of Airport Ground Handing services. Because of incidents, the duration of these services may sometimes take longer than expected. This may have consequences for preceding services or the planned takeoff of an aircraft. To avoid rescheduling of services, temporal buffers can be inserted between activities. The optimal buffer size depends on the occurrence of uncertain incidents. Stochastic approximation theory enables us to learn an optimal buffer based on observed incident costs. Convergence is however slow. The paper investigates the reason for the slow convergence and proposes an improvement that gives a speedup of a factor 30.