Multi-objective optimization of next-generation aircraft collision avoidance software

Abstract

Developed in the 1970's and 1980's, the Traffic Alert and Collision Avoidance System (TCAS) is the last safety net to prevent an aircraft mid-air collision. Although TCAS has been historically very effective, TCAS logic must adapt to meet the new challenges of our increasingly busy modern airspace. Numerous studies have shown that formulating collision avoidance as a partially-observable Markov decision process (POMDP) can dramatically increase system performance. However, the POMDP formulation relies on a number of design parameters modifying these parameters can dramatically alter system behavior. Prior work tunes these design parameters with respect to a single performance metric. This thesis extends existing work to handle more than one performance metric. We introduce an algorithm for preference elicitation that allows the designer to meaningfully define a utility function. We also discuss and implement a genetic algorithm that can perform multi-objective optimization directly. By appropriately applying these two methods, we show that we are able to tune the POMDP design parameters more effectively than existing work.