Automatic Hyperparameter Optimization in the Drone Racing Context

Abstract

Modern robotics is intertwined with artificial intelligence such as automatic controllers, and neural networks. Constructing such systems entail practitioners to carefully design hyperparameters of the intelligent modules embedded in the systems. Although such hyperparameters significantly affect the performance, finding an optimal configuration, in which the expert knowledge is required, becomes a tedious process. In this paper, we propose an evolutionary-based approach to perform automatic hyperparameter optimization, namely maximum velocity, acceleration and distance threshold at each gate, in the drone racing context. The hyperparameter update methods following the principal of optimality approach and greedy approach have been studied. The methods are evaluated in the high-fidelity drone racing simulator. The result suggests that the algorithm discovers the hyperparameter configuration that improves the best race time of the drone by a large margin from that of a human-designed configuration. © 2021, The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd.