Application of Evolutionary Optimization in Aircraft Collision Avoidance Algorithms

Н. В. Иванцевич, В. В. Худошин

Transactions of IAA RAS, issue 61, 53–58 (2022)

DOI: 10.32876/ApplAstron.61.53-58

Keywords: algorithm, collision avoidance, neural network, optimization, genetic algorithm, aircraft

Abstract

In the operational stage of an airborne collision avoidance system (CAS), there was generated a data base for further analysis by leading avionics experts. As a result of the analysis this international working group developed system improvement recommendations and issued a corresponding document containing the functional failure analysis. In particular, several types of air conflict situations were considered, where the algorithm would generate undesirable recommendations to the crew. One of the proposed problem solutions was algorithm modification, which would allow clarifying the generated recommendation. The purpose of the research was to study performance improvement possibilities of the existing algorithm by evolutionary optimization methods. The subject of the research was a previously modified algorithm, which was a combination of a standard working scheme and an additional algorithm, which refined the generated recommendation, based on a classic trained neural network. A sufficient amount of data obtained from the analysis of an air conflict situation was used to train the neural network. A deeper analysis of the modified algorithm run results revealed a possibility to increase qualitative characteristics. A suitable quality improvement method in this case was the application of evolutionary algorithms for neural network training to refine the connection weights. The carried-out research helped to review the application methods of evolutionary optimization of the neural network connection weights, which allowed to improve the determination quality of recommended manoeuvring procedures to the crew. The considered air situation type scenarios, which provide the improved algorithm training, were classified. A comparative analysis run results of the basic collision avoidance algorithm and the neural network before and after weights optimization was performed. An increased quality of the generated recommendations was recorded.

Citation