Mondrian Conformal Prediction of Boosting for Swarm Behavior Recognition

The relationships between empirical data of agents' movements and swarm behavior signals make it possible to predict swarm behaviors based on empirical data. From the theorem of view, the existing machine learning algorithms produce a function mapping the objects into scores and it is obviously practical demand for real-life data applications to provide uncertainty quantification for the simple prediction results. In this paper, we describe a distributions-free framework to measure the uncertainty of the simple prediction and we apply Mondrian Conformal Predictors to provide a valid set prediction under unconstrained randomness assumption. Both three representative behaviors of swarm real-life datasets, such as Align, Flock, and Group, and the synthesis collective swarm behavior datasets are presented, in addition, the validity check plots are also established for their prediction performance. We choose the modern machine learning algorithm, Boosting, as the role of underlying algorithms to produce the score of prediction rules. Instead of average treatment effect, the proposed method provides the two measurements, confidence, and credibility, for individual treatment effects and the set prediction has guaranteed average coverage in finite samples regardless of the unknown data distributions. Both Real-life and the synthesis datasets experimental results empirically demonstrate that existing methods suffer from a significant embarrassment in uncertainty quantification, while our methods achieve the desired and straightforward to provide the guaranteed probability.