“Combining q-learning and search with amortized value estimates,”