An objective penalty function of bilevel programming (Q430961)

The authors provide a penalty method for bilevel programming problems according to \[ (\text{BP}):\qquad\min f_1(x,y)\quad\text{s.t. }g_i(x,y)\leq 0,\quad i= 1,\dots, p,\qquad\text{and} \] \[ y\text{ solves the lower level problem }P(x):\min f_2(x,y)\text{ s.t. }h_j(x,y)\leq 0,\;j\in 1,\dots, q, \] where \(f_1,f_2,g_i,h_j: \mathbb{R}^n\times \mathbb{R}^m\to \mathbb{R}\) are given functions. Especially the functions \(f_i \) and \(h_j\) are assumed to be convex with respect to \(y\) i.e. the lower level problem is assumed to be convex. Using the functions \(P,Q:\mathbb{R}\to \mathbb{R}_+\cup\{\infty\}\) which are convex, differentiable, increasing for \(t> 0\) and vanishing for \(t= 0\), and using suitable (small) numbers \(M\) and \(N\), the penalty functions \[ \begin{aligned} F_1(x,y;M) &:= Q(f_1(x,y)- M)+ \sum P(g_i(x,y))+ \sum P(h_j(x,y)),\\ F_2(x,y;N) &:= Q(f_2(x,y)- N)+ \sum P(h_j(x,y))\end{aligned} \] and the following free optimization problem \[ \text{BP}(M,N)\;\min F_1(x,y;M)+\|\nabla y JF_2(x,y; N)\|^2\text{ s.t. }(x,y)\in\mathbb{R}^n\times \mathbb{R}^m \] are introduced. In the main theorems, relations are pointed out between the solutions of (BP) and the solutions of BP\((M,N)\) for suitable \(M\) and \(N\). Using these results, a so-called BPOPFA algorithm is developed to compute a globally optimal solution to (BP). The convergence of the algorithm is proved under weak assumptions.