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

This paper presents an element of an Artificial Intelligence (AI) system to assist operators with the management of traffic in orbit.
The system is based on a classification of collision events using DempsterShaffer’s theory of Evidence DSt). DSt is proposed to capture and quantify the epistemic uncertainty in the estimation of the Probability of Collision (PC). By capturing the epistemic uncertainty in the calculation of PC, we mitigate the paradoxical Dilution of Probability that affects the usual definition of this quantity. This phenomenon gives the counter intuitive idea that the lower the amount of information available to the operators, the smaller the probability of collision. This lack of information corresponds to an uncertainty that is epistemic in nature. Furthermore, when different sources provide contradictory information, the level of epistemic uncertainty in the calculation of PC can lead to a false confidence in the likelihood of a collision with either an undesirable increase in the number of collision avoidance manoeuvres or an equally undesirable number of false negative.
In order to make the classification automatic and provide continuous decision support to operators, we investigated the use of different Machine Learning techniques and identified the algorithm of choice to deliver the correct classification of collision events. We first considered a classification where the Belief and Plausibility in the correctness of the PC were used as additional classification criteria. Then we tested a second classification method that avoids the direct computation of Belief and Plausibility but retains the same added information on the credibility of the PC. Results suggest that Machine Learning can be effectively used in conjunction with DSt to provide decision support to operators and render the conjunction and collision analysis automatic.