On the use of machine learning and evidence theory to improve collision risk management

Sánchez, Luis and Vasile, Massimiliano and Minisci, Edmondo (2020) On the use of machine learning and evidence theory to improve collision risk management. In: 2nd IAA Conference in Space Situational Awareness, 2020-01-14 - 2020-01-16, Hilton Arlington.

[thumbnail of Sanchez-etal-ICSSA2020-On-the-use-of-machine-learning-and-evidence-theory-to-improve-collision-risk-management]
Text. Filename: Sanchez_etal_ICSSA2020_On_the_use_of_machine_learning_and_evidence_theory_to_improve_collision_risk_management.pdf
Accepted Author Manuscript

Download (607kB)| Preview


This paper introduces an Artificial Intelligence-based system to support operators to manage the risk of collision. The system is based on the concepts of Belief and Plausibility coming from Dempster-Shaffer's Evidence Theory applied to collision risk assessment. A revised calculation of the Probability of Collision, Pc) is proposed to mitigate the Dilution of Probability that affects the usual definition of this quantity. This phenomenon gives the counterintuitive idea that the lower the quality of the data (or amount of information available to the operators), the smaller the probability of collision. When different sources provide contradictory information, bigger uncertainties are considered which can lead to false confidence in the likelihood of a collision or forces operators to accept very large margins. The method presented here will account for epistemic uncertainty under the assumption of Evidence Theory which leads to the definition of confidence intervals on the probability of a collision. Confidence intervals incorporate the dependency of the probability of collision on the amount and quality of the available information, using the concepts of Belief and Plausibility introduced in Evidence Theory. The result of this revised calculation of the Pc is a more informed decision. At the same time, a lack of information can lead to a higher uncertainty on the decision to be made. Thus the paper will propose a possible approach to make optimal decisions under epistemic uncertainty, considering a given conjunction geometry and the time to the encounter. In addition to this new approach, an Artificial Intelligence-based system is applied to automatically provide the optimal decisions. A virtual database with a set of encounter geometries and associated uncertainties intervals have been created for training and validating the system. A set of Machine Learning techniques has been used to obtain preliminary results on the potential performance of the system. The system is presented under the form of a classification, where each of the classes for an encounter event is a suggested decision for the operator. Two approaches have been proposed. The first of them uses values of Belief and Plausibility at certain Pc thresholds and the time to the encounter for predicting the class. Very accurate results are provided by the techniques tested. The second approach uses the geometry of the encounter, allowing to skip the time-consuming step of computing Belief and Plausibility. Results suggest that Machine Learning techniques can be applied for obtaining an Artificial Intelligence-based system for supporting operators, although improvements on the methods should be done and a systematic analysis comparing techniques is recommended.