Prof. Neng-Fa Zhou (creator of Picat) mused that he’d retire once AI could build a CSP solver that outperforms his own — and argued that time hasn’t arrived: AI can augment the best human-built solvers and win at the portfolio/selection meta-game, but pure end-to-end learned solvers are still only competitive on narrow benchmarks. The real progress, he suggests, is in hybrid approaches — classical solver infrastructure paired with AI-learned policies.
Does anyone know of a project aiming for an AlphaGo- or AlphaFold-style breakthrough specifically for constraint satisfaction?