Scholtes relaxation method for pessimistic bilevel optimization

Abstract

https://www.southampton.ac.uk/maths/about/staff/abz1e14.pageWhen the lower-level optimal solution set-valued mapping of a bilevel optimization problem is not singlevalued, we are faced with an ill-posed problem, which gives rise to the optimistic and pessimistic bilevel optimization problems, as tractable algorithmic frameworks. However, solving the pessimistic bilevel optimization problem is far more challenging than the optimistic one; hence, the literature has mostly been dedicated to the latter class of the problem. The Scholtes relaxation has appeared to be one of the simplest and efficient way to solve the optimistic bilevel optimization problem in its Karush-Kuhn-Tucker (KKT) reformulation or the corresponding more general mathematical program with complementarity constraints (MPCC). Inspired by such a success, this paper studies the potential of the Scholtes relaxation in the context of the pessimistic bilevel optimization problem. To proceed, we consider a pessimistic bilevel optimization problem, where all the functions involved are at least continuously differentiable. Then assuming that the lower-level problem is convex, the KKT reformulation of the problem is considered under the Slater constraint qualification. Based on this KKT reformulation, we introduce the corresponding version of the Scholtes relaxation algorithm. We then construct theoretical results ensuring that a sequence of global/local optimal solutions (resp. stationarity points) of the aforementioned Scholtes relaxation converges to a global/local optimal solution (resp. stationarity point) of the KKT reformulation of the pessimistic bilevel optimization. The results are accompanied by technical results ensuring that the Scholtes relaxation algorithm is well-defined or the corresponding parametric optimization can easily be solved.

Value of distribution system information for DER deployment

“…The obtained solution is the strongest possible stationarity result (C-stationary point), owing to a very small value of 𝜌 used in the relaxation scheme [28]. The optimal solution of Scholtes's relaxation converges to the optimal solution of the KKT reformulation of the bilevel optimization, resulting in minimum analytical modeling errors [38].…”