Reproduce “Reinforcement Learning of Adaptive Acquisition Policies for Inverse Problems” on OpenResearch

About this paper

Gianluigi Silvestri, Fabio Valerio Massoli, Tribhuvanesh Orekondy, Afshin Abdi, Arash Behboodi

Researchers at OnePlanet Research Center, Qualcomm AI Research, and other institutions developed a generalized reinforcement learning framework for adaptive signal acquisition in inverse problems. This framework jointly optimizes both measurement policies and signal reconstruction, demonstrating improved recovery quality with fewer measurements, especially for low acquisition horizons and Radon transform measurements.

Key insights

• Jointly training the acquisition policy and the reconstruction model (end-to-end) is more effective than training them in stages or using pre-trained reconstruction models.
• Probabilistic adaptive schemes, coupled with a focus on average-case error, can bridge the gap between theoretical predictions and the observed practical benefits of adaptive sensing.
• A structured latent space derived from a variational formulation can aid the policy in making more effective decisions over extended acquisition trajectories, particularly in complex scenarios.

Summary by alphaXiv.