Gradient-Projection Methods


Active set methods have been criticized because the working set changes slowly; at each iteration at most one constraint is added to or dropped from the working set. If there are constraints active at the initial , but constraints active at the solution, then at least iterations are required for convergence. This property can be a serious disadvantage in large problems if the working set at the starting point is vastly different from the active set at the solution. Consequently, recent investigations have led to algorithms that allow the working set to undergo radical changes at each iteration and to interior-point algorithms that do not explicitly maintain a working set.

The gradient-projection algorithm is the prototypical method that allows large changes in the working set at each iteration. Given , this algorithm searches along the piecewise linear path

where is the projection onto the feasible set. A new point

is obtained when a suitable is found. For bound-constrained problems, the projection can be easily computed by setting

where is the middle (median) element of a set. The search for has to be done carefully since the function

is only piecewise differentiable.

If properly implemented, the gradient-projection method is guaranteed to identify the active set at a solution in a finite number of iterations. After it has identified the correct active set, the gradient-projection algorithm reduces to the steepest-descent algorithm on the subspace of free variables. As a result, this method is invariably used in conjunction with other methods with faster rates of convergence.

Trust-region algorithms can be extended to bound-constrained problems. The main difference between the unconstrained and the bound-constrained version is that we now require the step to be an approximate solution of the subproblem

where

An accurate solution to this subproblem is not necessary, at least on early iterations. Instead, we use the gradient-projection algorithm to predict a step (the Cauchy step) and then require merely that our step, , satisfies the constraints in the trust-region subproblem with . An approach along these lines is used by VE08 and PORT 3. In the bound-constrained code in LANCELOT the trust region is defined by the -norm and , yielding the equivalent subproblem

where is the vector of all ones.

The advantage of strategies that combine the gradient-projection method with trust-region methods is that the working set is allowed to change rapidly, and yet eventually settle into the working set for the solution. LANCELOT uses this approach, together with special data structures that exploit the (group) partially separable structure of , to solve large bound-constrained problems.


Up To:

Bound Constrained Optimization.


treesig.gif (5961 bytes)

[ OTC Home Page | NEOS Guide | NEOS Server | Optimization Tree ]


Updated 28 March 1996