and their main research objective is to develop intelligent DAA systems to allow small UAS to share the same airspace with manned aviation. 2 CivilEAirspace Avoid Mitigation Solution RiskE Assessment HazardE Identification Alerting boundaries Collision riskElevel Decision making Dectect Sensor Fusion Target TrackingE Sensor Measurement Errors Track establishing Data association UA Traffic Fig. 1 Block diagram for the proposed DAA system.PIC make final decisions on whether maneuvers are required to avoid potential conflicts in mid-air encounters. In the mitigation solution module, small UAS DAA mitigation guidance is usually provided to answer three core concerns: when to maneuver, how to maneuver, and how long for maneuvers, all of which are related to maneuvering timing and choices such as the turning direction, the turn rate, and the duration. A number of guidance methods have been reviewed and summarized for solving such questions in [3,4]. For instance, a predefined guidance method is designed based on predefined rules to determine escape trajectories. This is efficient in specific encounters, but in most cases, it is less effective and less optimal than an optimized guidance method [5]. As for the system response time, the predefined guidance method can provide an avoidance maneuvering solution immediately [4,5]. The optimized guidance method, on the other hand, usually requires extra computation time to search for the best solutions from all possible maneuvering options in mid-air encounters [6]. To overcome this drawback for real-time decision making, the TCAS selects the least-aggressive vertical maneuver within a limited set of potential climb or descent maneuvers that can provide adequate separation between aircraft during mid-air encounters [3,7]. However, this TCAS strategy cannot be directly adopted for new DAA guidance systems † since preferred horizontal maneuvers have many more maneuvering options (e.g., various turn rates and heading changes) than vertical maneuvers. For solving this challenge, a novel learning-based decision tree method is therefore proposed and designed in this paper to provide real-time DAA guidance without demanding extra computation time to search for the best solutions from all possible maneuvering options. This method is inspired by a Google artificial intelligence (AI) program, AlphaGo, which has recently mastered the complex ancient Chinese board game, "Go", defeating the best human "Go" game players in the last two years [9,10]. AlphaGo is not directly programmed to play "Go" games; instead, it is designed to learn how to play the game by a general purpose algorithm through analyzing millions of human expert-played "Go" games (supervised learning) and AlphaGo self-played "Go" games (reinforcement learning) [9]. By training through supervised learning and reinforcement learning, two knowledge-based statistical networks: the policy network (how to play the game in the next run) and the value network (how to evaluate the probability of winning the game with the current dec...