A Cyclic Hyper-parameter Selection Approach for Reinforcement Learning-based UAV Path Planning

Unmanned Aerial Vehicles (UAVs) offer new ways to fulfil a variety of urban transportation and service solutions. The ability to successfully plan and re-plan paths across a complex urban environment remains an unsolved significant problem. New Q-learning approaches have potential to address this problem, however they must first learn complex environment spaces. A traditional challenge within this field is the selection of suitable learning hyper-parameters that assist a Q-learning algorithm in achieving an optimal learning policy. It is known that testing and evaluating multiple hyper-parameter combinations is computationally expensive. Thus, this paper proposes a new method for hyper-parameter self-tuning, cyclically assigning hyper-parameters within a single learning process, eliminating the need to experimentally seek optimal hyper-parameter value combinations. Evaluation of the captured results show, training with cyclical hyper-parameter exploration instead of fixed values, achieves improved path generation, while reducing the cumulative learning time required. Although the focus of this approach is centred around a Multi Q-table Path Planning solution, this work presents a practical tool applicable to Reinforcement Learning techniques generally.