Nicholas (Glenbrook South) & Anne (Naperville North)
Mentor: Brian Nord
Reinforcement learning is a method of machine learning that is most suitable for skill-oriented tasks‒from playing chess to developing robotic fine-motor skills. This summer, we explored the applications of reinforcement learning in astronomical surveying strategy. Until now, most survey strategies for astronomical projects have been hand-developed by teams of scientists, as for the Sloan Digital Sky Survey, or tuned through human operated computer simulations, as for the Dark Energy Survey. The purpose of our project is to model astronomical data and implement a reinforcement learning algorithm that could optimize an observation schedule for the night sky.
The reinforcement learning program receives data representing states of celestial objects and optimizes over time allocation for each object. The software design primarily utilizes neural networks to serve as a foundation for reinforcement learning. The neural networks are created using Google’s open-source library, TensorFlow, and trained iteratively using principles of reinforcement learning. Reward functions, as defined by the user, are adjusted and fed to the neural network to define the objectives of the optimization problem. Using sample data from Astroplan’s python library, we implemented a neural network with reinforcement learning that generates a near-optimal schedule for a given location and night. The project’s next steps would be to implement a neural network design that could perform image processing and compute on larger datasets to better represent the telescope-sky system. The ultimate aim of our project is to incorporate reinforcement learning into the development of surveying strategies for large astronomical projects.