Resumen
To assist air traffic controllers (ATCOs) in resolving tactical conflicts, this paper proposes a conflict detection and resolution mechanism for handling continuous traffic flow by adopting finite discrete actions to resolve conflicts. The tactical conflict solver (TCS) was developed based on deep reinforcement learning (DRL) to train a TCS agent with the actor?critic using a Kronecker-factored trust region. The agent?s actions are determined by the ATCOs? instructions, such as altitude, speed, and heading adjustments. The reward function is designed in accordance with air traffic control regulations. Considering the uncertainty in a real-life situation, this study characterised the deviation of the aircraft?s estimated position to improve the feasibility of conflict resolution schemes. A DRL environment was developed with the actual airspace structure and traffic density of the air traffic operation simulation system. Results show that for 1000 test samples, the trained TCS could resolve 87.1% of the samples. The conflict resolution rate decreased slightly to 81.2% when the airspace density was increased by a factor of 1.4. This research can be applied to intelligent decision-making systems for air traffic control.