Aircraft collision avoidance using Monte Carlo Real-Time Belief Space Search

Abstract

This thesis presents the Monte Carlo Real-Time Belief Space Search (MC-RTBSS) algorithm, a novel, online planning algorithm for partially observable Markov decision processes (POMDPs). MC-RTBSS combines a sample-based belief state representation with a branch and bound pruning method to search through the belief space for the optimal policy. The algorithm is applied to the problem of aircraft collision avoidance and its performance is compared to the Trac Alert and Collision Avoidance System (TCAS) in simulated encounter scenarios. The simulations are generated using an encounter model formulated as a dynamic Bayesian network that is based on radar feeds covering U.S. airspace. MC-RTBSS leverages statistical information from the airspace model to predict future intruder behavior and inform its maneuvers. Use of the POMDP formulation permits the inclusion of different sensor suites and aircraft dynamic models. The behavior of MC-RTBSS is demonstrated using encounters generated from an airspace model and comparing the results to TCAS simulation results. In the simulations, both MC-RTBSS and TCAS measure intruder range, bearing, and relative altitude with the same noise parameters. Increasing the penalty of a Near Mid-Air Collision (NMAC) in the MC-RTBSS reward function reduces the number of NMACs, although the algorithm is limited by the number of particles used for belief state projections. Increasing the number of particles and observations used during belief state projection increases performance.

(cont.) Increasing these parameter values also increases computation time, which needs to be mitigated using a more efficient implementation of MC-RTBSS to permit real-time use.