The quadratic balanced optimization problem

We introduce the quadratic balanced optimization problem (QBOP) which can be used to model equitable distribution of resources with pairwise interaction. QBOP is strongly NP-hard even if the family of feasible solutions has a very simple structure. Several general purpose exact and heuristic algorithms are presented. Results of extensive computational experiments are reported using randomly generated quadratic knapsack problems as the test bed. These results illustrate the efficacy of our exact and heuristic algorithms. We also show that when the cost matrix is specially structured, QBOP can be solved as a sequence of linear balanced optimization problems. As a consequence, we have several polynomially solvable cases of QBOP.

💡 Research Summary

The paper introduces a novel combinatorial optimization model called the Quadratic Balanced Optimization Problem (QBOP). In QBOP, a subset of items must be selected under conventional feasibility constraints (such as cardinality, knapsack capacity, or group restrictions), and the objective is to minimize the “balance gap” – the difference between the largest and the smallest pairwise interaction costs among the selected items. Formally, given a cost matrix C = (c_{ij}) and a feasible family ℱ ⊆ 2^N, the goal is to find S ∈ ℱ that minimizes max_{i,j∈S} c_{ij} – min_{i,j∈S} c_{ij}. This formulation captures situations where fairness or equity is required not only at the level of individual items but also with respect to their mutual interactions (e.g., synergy, conflict, or interference).

The authors first establish the computational hardness of QBOP. By a reduction from the classic Partition and Quadratic Knapsack problems, they prove that QBOP is strongly NP‑hard even when ℱ has a very simple structure (e.g., all subsets of size at most k or all knapsack‑feasible subsets). Consequently, the presence of quadratic terms dramatically increases the problem’s complexity compared to linear balanced optimization, which is often polynomially solvable.

To address this difficulty, the paper proposes several exact solution strategies. A branch‑and‑bound framework is built around a sequence of linear balanced optimization (LBO) subproblems that provide lower bounds; each node’s bound is obtained by solving an LBO that fixes a subset of decisions and relaxes the rest. Additionally, a mixed‑integer linear programming (MILP) formulation introduces auxiliary variables M and m to represent the maximum and minimum pairwise costs, respectively, and minimizes M – m subject to standard feasibility constraints. The MILP can be handed to commercial solvers (CPLEX, Gurobi) and yields optimal solutions for modest‑size instances.

Given the intractability for larger instances, the authors design a suite of heuristics. A constructive greedy algorithm builds an initial feasible set by iteratively adding items that cause the smallest increase in the balance gap. This solution is then refined by a local‑search procedure that swaps, inserts, or deletes items to further reduce the gap. To enhance global exploration, the paper integrates meta‑heuristics such as simulated annealing, tabu search, and a genetic algorithm, forming a hybrid framework that balances intensification and diversification.

A particularly important contribution is the identification of polynomially solvable special cases. When the cost matrix is monotone (i.e., c_{ij} ≤ c_{i+1,j+1} for all i < j) or block‑diagonal (the matrix decomposes into independent sub‑matrices), the QBOP can be reduced to a sequence of linear balanced problems. In these scenarios, the balance gap can be decreased stepwise by solving a series of LBOs, each of which is solvable in polynomial time. The overall complexity becomes O(n·poly(m)) where n is the number of items and m the number of blocks, providing efficient exact algorithms for a broad class of practical instances.

The experimental section evaluates the proposed methods on randomly generated quadratic knapsack instances. Problem sizes range from 50 to 200 items, with capacity ratios between 0.2 and 0.8, and cost matrices drawn from uniform, normal, and sparse distributions. Exact methods (branch‑and‑bound and MILP) solve all instances with n ≤ 80 within one hour, but experience exponential growth for larger n. The greedy‑plus‑local‑search heuristic attains solutions within 2–5 % of the optimum in under one second on average, while the hybrid meta‑heuristic improves the gap to 1–2 % at the cost of 5–10 seconds of runtime. For the structured matrices (monotone or block‑diagonal), the polynomial algorithm solves instances with n = 200 in a few milliseconds, matching the optimal value. These results demonstrate that, despite the theoretical hardness, QBOP can be tackled effectively in practice, especially when problem structure is exploited.

In conclusion, the paper establishes QBOP as a meaningful extension of balanced optimization that captures pairwise equity considerations, proves its strong NP‑hardness, and supplies a comprehensive algorithmic toolbox—including exact, heuristic, and structure‑exploiting methods. Future research directions suggested include extending polynomial algorithms to broader classes of sparse or hierarchical cost matrices, developing online or dynamic versions of QBOP for real‑time applications, and integrating QBOP with other combinatorial problems (e.g., multi‑objective scheduling, network design) to model more complex decision‑making environments.