Reinforcement Learning for Data-Driven Workflows in Radio Interferometry. I. Principal Demonstration in Calibration

TL;DR

This work tackles the automation of data reduction in radio interferometry by framing calibration and flagging as actions in a reinforcement learning environment. Using simulated VLA-like data, it defines a compact state space, six global actions, and a joint objective that balances residual Gaussian-approximation quality with runtime, via $action\text{-}value = 1\times 10^{6}\cdot EMD + runtime$. The results show that Q-learning can discover optimal single actions and, with RFI, optimal action sequences; neural networks and decision trees provide continuous and interpretable mappings from data state to actions. The study demonstrates data-driven automation, heuristic discovery, and tool diagnostics, and outlines a clear path toward applying these ideas to real data and including imaging stages for self-calibration and scalable automation.

Abstract

Radio interferometry is an observational technique used to study astrophysical phenomena. Data gathered by an interferometer requires substantial processing before astronomers can extract the scientific information from it. Data processing consists of a sequence of calibration and analysis procedures where choices must be made about the sequence of procedures as well as the specific configuration of the procedure itself. These choices are typically based on a combination of measurable data characteristics, an understanding of the instrument itself, an appreciation of the trade-offs between compute cost and accuracy, and a learned understanding of what is considered "best practice". A metric of absolute correctness is not always available and validity is often subject to human judgment. The underlying principles and software configurations to discern a reasonable workflow for a given dataset is the subject of training workshops for students and scientists. Our goal is to use objective metrics that quantify best practice, and numerically map out the decision space with respect to our metrics. With these objective metrics we demonstrate an automated, data-driven, decision system that is capable of sequencing the optimal action(s) for processing interferometric data. This paper introduces a simplified description of the principles behind interferometry and the procedures required for data processing. We highlight the issues with current automation approaches and propose our ideas for solving these bottlenecks. A prototype is demonstrated and the results are discussed.

Figures (13)

• Figure 1: Highlighting the possibility of loops in processing steps (dotted lines) rather than a linear flow (solid lines).
• Figure 2: Visibility amplitudes plotted in 3D for one baseline. Other baselines in the dataset are similar in shape but vary in intensity according to the antenna-to-antenna variations. X-Y axes form the time-frequency plane with visibility amplitude on the z-axis (also indicated by color). Left: A simulation without any gain distortions, which is a flat time-frequency plane with noise. Center: Visibilities containing a large gain distortion that varies along the time axis that would require calibration corrections for each timestep. This simulation would benefit from averaging along frequency prior to calibration. Right: A simulation dominated by an RFI outlier that would need to be removed before estimating any potential underlying gain distortions.
• Figure 3: Figure: An example comparison between theoretical noise distribution (green) and a model’s residuals distribution (red) after averaging both axes before calibration. The faded magenta and green distributions are the original sampled values while the solid magenta and green lines represent the distributions after the Savitzky-Golay filter is applied. The blue highlights the difference between the smoothed distributions from which the EMD is calculated, indicating correctness (lower is better). The vertical dashed line indicates where the 0 point is; theoretical noise is centered around 0.
• Figure 4: Leftmost: Plotted visibilities of a simulated dataset containing gain distortions. Branching from that simulation is each possible action and the results each action has on the data. Each row of images associated with each action (from L→ R) are the resulting time-frequency plane, the image residuals, and the residuals distribution compared to the theoretical noise distribution. Color scales and X-Y axes values are synchronized on all plots to make relative visual comparison easy. Image residuals that look noise-like, and residual distributions that look Gaussian, indicate better numerical outcomes such as row 1 and 4. Rows 1 and 4 have RMS levels of 1e-5Jy/beam whereas rows 2, 3, 5, and 6 have an average of 3e-2Jy/beam.
• Figure 5: 3D state-space with thousands of simulations colorized by best the performing action. State space axes are frequency variation, time variation, and mean gain. The individual points are simulations with those statistical properties. The points are colorized by the best performing action, as determined by our metric. We can see regions in the state space that correspond to ideal actions. Flag action will come later.