ALM method may be called as Method of Multiplier (MOM) or Primal-Dual Method. Let’s consider Lagrangian functional only for equality constraints.
Now, for a Lagrange multiplier vector , suppose that there is an optimum for the following unconstrained optimization problem.
If satisfy all the equality constraints in the original design problem, is an optimum for the original optimization problem and is a Lagrange multiplier optimum. Consequently, the original optimization problem can be transformed into the following problem that have the same optimum and .
subject to
In order to avoid the unboundness of Lagrangian, a penalty function is introduced. We call it as augmented Lagrangian.
where, is the penalty parameter for the ith equality constraint. In the ALM method, the unconstrained optimization tool sequentially minimize the augmented Lagrangian for the given value of and . Then, these two parameters are modified to satisfy the optimality condition.
The update rule for Lagrange multipliers can be determined from the following relation.
This implies
Hence, the update rule for Lagrange multipliers is
where, the superscript is the iteration of ALM algorithm.
Inequality constraints are transformed into equality constraints by adding slack variables . Thus, the augmented Lagrangian becomes
.
Then, a new primal variables are . The augmented Lagrangian should satisfy the optimality condition for slack variable .
Hence, an optimum of slack variable is
.
Now, these optimum values are substituted into the originally transformed form.
Hence, the augmented Lagrangian for inequality constraints are transformed into the following simple functional.
where, . Also, the Lagrange multiplier update rule is defined as
.
In order to sequentially solve the augmented Lagrangian, AutoDesign uses quasi-Newton(BFGS) for and conjugate gradient method (Hestnes-Stiefel) for .