IST Lunch Bunch

The process of scientific discovery is based on intuition, trial and error and oftentimes plain guesswork.  Can this process be, to some degree, turned into an algorithm?  Can discovery be computed?  Although these questions seem unorthodox their answer appears to be positive for the discovery of scalable numerical solvers where discovery can be automated by reformulating the process of computing with partial information and limited resources as that of playing underlying hierarchies of adversarial information games.  As an illustration we show how the application of the proposed approach leads to the discovery of scalable multigrid/multiresolution algorithms (of lowest known complexity with rigorous a-priori accuracy and performance estimates) for PDEs with rough coefficients.