Automated Data-Driven Decision Making for Radio Calibration and Imaging

Data taken with a radio interferometer must pass through a sequence of calibration and imaging stages in order to produce images that an astronomer can use to study astrophysical phenomena. These processing stages include data editing, calibration of the instrument hardware, and image reconstruction through deconvolution. Steps within each of these stages may be assembled in a variety of sequences and combinations, referred to as ‘recipes’. Any single end-to-end data analysis recipe may be thought of as one possible path through all possible configurations. The full configuration space of the software tools is quite vast.

Current automated pipelines essentially hard-code a “few” most-used paths through this entire maze using a small set of heuristics-based decision points. While these recipes do process ~80% of the data volume, they have yet to achieve true science-ready results for most projects (i.e. requiring no further manual tuning and which may directly be used for astrophysical interpretation). Most of the current pipeline results need to be reviewed by a Data Analyst for quality assessment for two reasons: 1) there is often no objective set of metrics to define correctness that applies in all situations and 2) errors or artifacts that surface during processing are not handled by the recipes. Therefore, Data Analysts may need to optimize or reprocess individual steps or entire sequences after the pipeline has executed before distributing data.

Our research is developing a system that decides the optimal action at every step of processing based on the state of the input data. This requires a software framework where actions can be evaluated and estimated within context. This research has started with building the fundamental components required to apply Deep Reinforcement Learning. We have identified a simplified scenario that exhibits the larger problems we want to address and have begun to test the applicability of the new methodologies to this scenario while keeping interpretation of results clear. We will report on our progress and research roadmap.