A multi-agent reinforcement learning approach to autonomous aircraft taxiing with taxiing time, fuel consumption, and emission optimization

Traditional research on airport surface movement largely relies on heuristic algorithms for aircraft routing and speed planning. These algorithms require extensive computation time and lack dynamic decision-making based on real-time taxiway traffic, limiting their ability to support efficient airport taxiing and potentially leading to conflicts, additional fuel consumption, and increased emissions. Therefore, this study proposes a novel approach that consists of a multi-agent reinforcement learning (MARL) algorithm and a Graph Attention Network (GAT) to train a model for aircraft taxiing control. The GAT leverages airport taxiing data in a graph format, and MARL is a promising data-driven approach for training a neural network for airport taxiing control modeling. This innovative method processes data through a multi-layer graph neural network, subsequently channeling it into a deep reinforcement learning (DRL) framework with a residual network component, designed specifically for making strategic taxiing decisions. The algorithm's performance is refined via a reward system that focuses on minimizing taxiing time, fuel consumption, emission and conflict to ensure efficient taxiing. Once trained, the MARL model is capable of autonomously generating optimal and conflict-free taxiing control strategies based on real-time taxiway traffic conditions. Our proposed method demonstrates the efficacy and adaptability of a 1-h airport flight plan at an international airport. The results show that our MARL approach significantly outperforms traditional reinforcement learning methods and the conventional shortest path algorithm by optimizing taxiing time, fuel consumption, and emission.