Reproduce “Seasonal Station-Keeping of Short Duration High Altitude Balloons using Deep Reinforcement Learning” on OpenResearch

About this paper

Tristan K. Schuler, Chinthan Prasad, Georgiy Kiselev, Donald Sofge

Researchers at the U.S. Naval Research Laboratory developed a deep reinforcement learning strategy for autonomous station-keeping of short-duration high-altitude balloons. This involved creating a synthetic wind model to introduce realistic atmospheric uncertainties into simulations and a "Forecast Score" algorithm to assess mission viability based on wind diversity, with DRL agents achieving around 50% time within a 50 km radius.

Key insights

• Deep Reinforcement Learning, specifically DQN, can effectively learn and execute station-keeping strategies for short-duration solar high-altitude balloons by exploiting vertical wind diversity.
• Training DRL agents with synthetic wind forecasts that simulate realistic deviations from standard reanalysis data is crucial for developing robust policies capable of handling real-world atmospheric uncertainties.
• The degree of vertical wind diversity, as quantified by the "Forecast Score" algorithm, is a strong indicator of the potential for successful HAB station-keeping and exhibits significant seasonal and regional variations.

Summary by alphaXiv.