Description
The New Robotic Telescope (NRT) represents a significant advancement in autonomous astronomical observation, with core science goals requiring response times under 30 seconds for transient events. Traditional scheduling approaches often rely on fixed priority rankings that inadequately address time-sensitive observations, especially as astronomical surveys like LSST generate unprecedented volumes of alerts requiring rapid follow-up.
We present a novel scheduling framework that employs Deep Q-network reinforcement learning to optimize the critical trade-off between observation priority and response time. Our research demonstrates that conventional priority-based scheduling can be significantly enhanced by incorporating a response time parameter within the reward function. Through comparative analysis of different reward weight configurations, we identified a non-linear relationship where a 1:2 priority-to-response-time weighting achieves optimal performance - maintaining traditional priority distribution while dramatically improving response time success rates from ~12% to ~82%.
This research addresses the computational challenges of modern time-domain astronomy by developing an adaptive, learning-based scheduler that can process large alert streams efficiently. Our implementation demonstrates how machine learning techniques can be leveraged to maximize scientific return in resource-constrained environments, particularly when competing objectives must be balanced. The scalable approach makes it suitable for next-generation observational facilities producing massive data volumes requiring real-time decision-making.
