PROJECT TITLE :

Physics Informed Deep Reinforcement Learning for Aircraft Conflict Resolution

ABSTRACT:

It is proposed that physics-informed Deep Reinforcement Learning (RL) can be used to devise an innovative approach to the resolution of aircraft conflicts in air traffic management (ATM). The reason for this is to incorporate previous physics knowledge and a model into the learning algorithm in order to simplify the process of finding the best policy and to present results that are understandable to humans for the purposes of display and decision-making. To begin, the information regarding the quantity of intruders as well as their speeds, heading angles, and positions are integrated into an image by using the solution space diagram (SSD), which is utilized in the ATM for conflict detection and mitigation. The prior physics knowledge from the ATM domain that serves as the input features for learning is provided by the SSD. In order to carry out the deep reinforcement learning, a Convolutional Neural Network is combined with the SSD images. The next step is to develop a network of actors and critics in order to study conflict resolution policy. The proposed methodology is demonstrated with the help of several different numerical examples. The idea of physics-informed learning, as it is currently conceived, is applied to investigation of both discrete and continuous RL. There is an in-depth comparison made between the proposed algorithm and traditional RL-based conflict resolution, as well as a discussion of both. The method that has been suggested can manage an arbitrary number of intruders, and it also demonstrates faster convergence behavior thanks to the encoded prior physics understanding. The learned optimal policy is also beneficial for proper display to support decision-making, which is something that is supported by this. Based on the findings of the current investigation, a number of significant conclusions and recommendations for further research are presented.