|
Success stories 2012
Francqui Chair (ULg 2011-2012) awarded to Yurii Nesterov : New birth of Fast Gradient Methods ![]() Gradient methods were among the first schemes developed for nonlinear optimization problems. Their main advantage is a very cheap iteration cost. However, their rate of convergence is often slow. This is especially true for nonsmooth optimization problems, where these methods need hundreds of thousands of iterations for obtaining a reasonable approximation to the optimal solution. At the same time, it was proved by Complexity Theory, that their rate of convergence is the best possible one.
During last twenty years, these methods were almost out of the serious computational practice. The situation was changed in 2005, when Yurii Nesterov discovered a way for significant acceleration of the gradient schemes. His approach is based on artificial smoothing of the objective function by an appropriate modification of the initial problem. Then the new problem can be solved by Fast Gradient Methods, the known technique developed in early 80s, which was not very popular among practitioners because of the absence of interesting applications. As a result, the number of iterations of new methods becomes proportional to a square root of the number of iterations of the old ones. When this number is of the order of millions, the gain is very substantial, taking into account that the cost of one iteration of the new methods is the same as before.
Now we can see an explosion of interest to Fast Gradient Methods. They find more and more applications in Image Processing, Pattern Recognition, Statistics, and many other fields. The corresponding theoretical results were the main motivation for awarding to Yurii Nesterov Chaire Francqui 2012 by Liege University.
![]() Vincent Blondel, Petar Kokotovic Distinguished Visiting Professor
|
12/04/2013
|
|
||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||