A Dual Recurrent Neural Network-based Hybrid Approach for Solving Convex Quadratic Bi-Level Programming Problem

The current paper presents a neural network-based hybrid strategy that combines a Genetic Algorithm (GA) and a Dual Recurrent Neural Network (DRNN) for efficiently and accurately solving the quadratic-Bi-level Programming Problem (BLPP). In this model, the GA is used to handle the upper-level decision problem by choosing desirable solution candidates and passing them to the lower-level problem. Subsequently, in the lower-level, the parameterized-DRNN is used to determine possible optimal solutions. This combination offers several benefits such as being a parallel computing structure, the RNN offers faster convergence to the optimum for the lower-level decision problem and it also helps to quickly and accurately determining the global optimal. Moreover, the GA can quickly reach the global optima and can search without becoming stuck to the local optimal. Additionally, by choosing desirable initialization of parameters, the proposed algorithm reaches the optimum with higher accuracy. Apart from that, there are still a few utilizations of hybrid NN-based methods for solving BLPPs. Hence, we believe the proposed algorithm will contribute to solving quadratic-BLPPs involved in various engineering, management, and finance applications. The accuracy and efficiency of the proposed method have been found better than the existing and widely used approaches, while doing experimental verification using four well-known examples used in prior works. © 2020 Elsevier B.V.