• offline dataset
• deep learning
• schedule optimization
• artificial intelligence
• self-driving telescope
• reinforcement learning

Scientists used to manually design the schedule of observatories, considering a list of astronomical objects to visit, and gathering data from, while dealing with several external factors.
In this thesis work, the aim is the exploration of how artificial intelligence, and in particular Reinforcement Learning (RL), can optimize the data-gathering process itself, maximizing the potential of telescopes as powerful tools able to unveil the mysteries of the cosmos.
Familiar RL algorithms and some of their adaptations have been evaluated using an offline dataset, containing interactions between a telescope and a portion of the celestial space. The dataset contained a discrete range of locations in the sky that the telescope was supposed to visit.
The primary focus was on evaluating RL algorithms that are specifically designed to accommodate discrete action spaces, conducting comparisons between policy-based, value-based methods, and methods in the class of evolutionary computation.
This thesis work also resulted in a software framework that can be used to train and evaluate RL algorithms on any dataset containing data regarding interactions between a telescope and the sky.
Some advanced investigations on the problem are also considered, such as observation space reduction, reward modification, and interpretability.