固定点迭代的伴随方法.pdfVIP

14th European Workshop on AD ADJOINTS OF FIXED-POINT ITERATIONS Ala Taftaf INRIA, SOPHIA ANTIPOLIS, FRANCE Contexte général Adjoint Algorithms Parameters Adjoint Algorithm Adjoint (AD) Algorithm Gradient • Adjoint algorithms : the most efficient way to obtain gradient of a numerical simulation • Computing the gradients has a cost (time, memory) • One way around : take advantage of the structures of the given program (parallel loops, fixed point methods,..) Contexte général Fixed Point Iterations • Many implicit functions , = are computed with convergent iterations: Initial guess Iterate + () = () , To fixed point ∗ () = ∗ () , • Followed by an objective function = , 2 ∗ Contexte général Adjoint of Fixed Point Iterations • First iterations of a fixed-point search : = () , ….. operate on a meaningless state vector =No adjoint needed • Adjoin only the (few) last iterations • Save trajectory storage At least two authors have studied mathematically fixed point iterations with the goal of defining an efficient adjoint : Christianson and Griewank Contexte général Christianson’s strategy (BC) + = () , ∗ Forward Sweep = , ∗ = = . ( , ) ∗