On a Stackelberg Subset Sum Game

TL;DR

This paper studies a two-level Stackelberg game based on the subset sum problem, where a leader manipulates item weights to influence a follower's heuristic solution, analyzing complexity and providing algorithms.

Contribution

It introduces a novel Stackelberg subset sum game model with weight control, analyzing its complexity and proposing exact pseudopolynomial algorithms.

Findings

Optimal weight setting is NP-hard.

Constant-factor approximation is NP-hard.

Exact dynamic programming algorithms are developed.

Abstract

This contribution deals with a two-level discrete decision problem, a so-called Stackelberg strategic game: A Subset Sum setting is addressed with a set $N$ of items with given integer weights. One distinguished player, the leader, may alter the weights of the items in a given subset $L\subset N$, and a second player, the follower, selects a solution $A\subseteq N$ in order to utilize a bounded resource in the best possible way. Finally, the leader receives a payoff from those items of its subset $L$ that were included in the overall solution $A$, chosen by the follower. We assume that the follower applies a publicly known, simple, heuristic algorithm to determine its solution set, which avoids having to solve NP-hard problems. Two variants of the problem are considered, depending on whether the leader is able to control (i.e., change) the weights of its items (i) in the objective function or (ii) in the bounded resource constraint. The leader's objective is the maximization of the overall weight reduction, for the first variant, or the maximization of the weight increase for the latter one. In both variants there is a trade-off for each item between the contribution value to the leader's objective and the chance of being included in the follower's solution set. We analyze the leader's pricing problem for a natural greedy strategy of the follower and discuss the complexity of the corresponding problems. We show that setting the optimal weight values for the leader is, in general, NP-hard. It is even NP-hard to provide a solution within a constant factor of the best possible solution. Exact algorithms, based on dynamic programming and running in pseudopolynomial time, are provided. The additional cases, in which the follower faces a continuous (linear relaxation) version of the above problems, are shown to be straightforward to solve.