Multi-step Bayesian Optimization-based intelligent task planning for an ozone observation satellite

Abstract

This paper presents an intelligent task planning strategy for an Earth Observing Satellite (EOS) monitoring the Ozone distribution using multi-step Bayesian Optimization (BO). Significant improvement in observation performance can be achieved by implementing an offset-pointing strategy using single-step or classical BO compared to the simple nadir-pointing approach, which does not consider the possible future outcomes when deciding the current action. To account for future considerations, this paper presents a modified Continuous Belief Tree Search (CBTS) algorithm that enables a non-myopic multi-step look-ahead optimization in the continuous action domain. The design utilizes a Control-effort based Upper Confidence Bound function as the acquisition function for BO. Simulation results demonstrate that the multi-step BO is capable of significantly reducing the model uncertainty compared to the single-step BO. Additionally, using BO provides a superior way to determine the first actions than obtaining them randomly. Moreover, sampling the training data set provides a way to reduce the computational time without reducing the computation tree size.