Collision Avoidance Using Neural Networks Learned by Genetic Algorithms

As Air Traac keeps increasing, many research programs focus on collision avoidance techniques. In this paper, a neural network learned by genetic algorithm is introduced to solve connicts between two aircraft. The learned NN is then tested on diierent connicts and compared to the optimal solution. Results are very promising. 1 Air Traac Control and Collision Avoidance As Air Traac keeps increasing, overloading of the ATC 1 system becomes a serious concern. For the last twenty years, diierent approaches have been tried, and diierent solutions have been proposed. To be short, all theses solutions fall in the range delimited by the two following extreme positions: On the one hand, it could be possible to imagine an ATC system where every trajectory would be planned and where each aircraft would follow its trajectory with a perfect accuracy. With such a system, no reac-tive system would be needed, as no connict 2 between aircraft would ever occur. This solution is close to the Centre d'Etudes de la Navigation A erienne y Ecole Nationale de L'Aviation Civile z Ecole Nationale Sup erieure d'Electronique, d'Electro-technique, d'Informatique et d'Hydraulique de Toulouse 1 Air Traac Control 2 2 aircraft are said to be in connict if their altitude diier-ence is less than 1000 feet (305 meters) and the horizontal distance between them is less than 8 nautical miles (14800 meters). These two distances are respectively called vertical and horizontal standard separation ARC-2000 hypothesis, which has been investigated by the Eurocontrol Experimental Center 5]. On the other hand, it could also be possible to imagine an ATC system where no trajectories are planned. Each aircraft would ight its own way, and all collisions would have to be avoided by reactive systems. Each aircraft would be in charge of its own security. This could be called a completely free ight system. The free ight hypothesis is currently seriously considered for all aircraft ying \high enough" in a quite near future. Of course, no ATC system will ever totally rely on only one of these two hypothesis. It is quite easy to understand why. A completely planned ATC is impossible, as no one can guarantee that each and every trajectory would be perfectly followed; there are too many parameters that can not be perfectly controlled: meteorological conditions (storms, winds, etc.), but also breakdowns in aircraft (motor, aps, etc) or other problems (closing of landing runaway on airports, etc.). On …