Using Neuro-Evolution in Aircraft Deicing Scheduling

Resource scheduling problems with incomplete information and environment changes have been studied for decades. The vast majority of the research efforts in the scheduling problems under uncertainty assume a central authority with a global objective of maximizing the resource usage. In real-life scheduling problems, in addition to the dynamic environment, sometimes conflicting interests of different parties renders a centralized approach undesirable. Applying multi-agent approaches in resource scheduling overcomes the restrictions associated with traditional static centralized scheduling. This paper studies a multi-agent scheduling system in the context of an interesting airport planning problem: the planning and scheduling of deicing and anti-icing activities. In this application domain, self-interested aircraft agents have an incentive to reserve a deicing resource as early as possible, leading to sub-optimal schedules. To counter this effect, we propose the use of decommitment penalties, forcing agents to reserve the deicing resources at a later time point, which results in a better overall schedule. This paper investigates the effects of agents learning an ‘optimal’ strategy in this context. To learn an ‘optimal’ strategy, we apply genetic algorithms to train a neural network, the agent use to decide when to reserve the deicing resource. Experiments show that the neural-evolution algorithm outperforms the derived strategy based on simplified cost estimation in the decommitment penalties mechanism, which in turn significantly improve the efficiency and fairness compared with naive First Come, First Served approach.