Abstract:
The Federal Air Marshal Service provides front-line security in homeland defense by protecting civil aviation from potential terrorist attacks. Unique challenges arise in maximizing effective deployment of a limited number of air marshals to cover the risk posed by potential terrorists on nearly 30,000 daily domestic and international flights. Some risk presents in a stochastic nature (e.g., a last minute ticket sale where suspicion is aroused). Pre-scheduled air marshal deployments cannot respond to risk which presents stochastically in real-time. This dissertation proposes the formation of a quick reaction force to explicitly address stochastic risk of terrorism on commercial flights and presents a method for near real-time force allocation to optimize risk coverage. The dynamic allocation of reactionary air marshals requires sequential decision making under uncertainty with limited lead time. This dissertation investigates the application of an approximate dynamic program (ADP) to assist schedulers allocating air marshals in near real-time. ADP is a form of reinforced learning that seeks optimal decisions by incorporating future impacts rather than optimizing only on short-term rewards. The marshal allocation system is modeled as a Markov decision process. Due to the many variables and environment complexity, explicit storage of all states and their values is not possible. Value function approximation schemes are explored to mitigate scalability challenges by alleviating the need for state value storage. The study demonstrates that air marshal allocation in near real-time is possible using an ADP with value function approximation and results in improved coverage of stochastic risk over the myopic approach or pre-scheduling.
