On exact solution approaches for bilevel quadratic 0–1 knapsack problem

We consider the bilevel quadratic knapsack problem (BQKP) that model settings where a leader appropriates a budget for a follower, who solves a quadratic 0–1 knapsack problem. BQKP generalizes the bilevel knapsack problem introduced by Dempe and Richter (Cent Eur J Oper Res 8(2):93–107, 2000), where the follower solves a linear 0–1 knapsack problem. We present an exact-solution approach for BQKP based on extensions of dynamic programs for QKP bounds and the branch-and-backtrack algorithm. We compare our approach against a two-phase method computed using an optimization solver in both phases: it first computes the follower’s value function for all feasible leader’s decisions, and then solves a single-level, value-function reformulation of BQKP as a mixed-integer quadratically constrained program. Our computational experiments show that our approach for solving BQKP outperforms the two-phase method computed by a commercial state-of-the-art optimization software package.