[Paper Review] Evaluation and Improvement of Laruelle-Widgrén Inverse Banzhaf Approximation

Frits de Nijs, Daan Wilmer|arXiv (Cornell University)|Jun 6, 2012

Game Theory and Voting Systems9 references3 citations

TL;DR

This paper evaluates and improves the Laruelle-Widgrén inverse Banzhaf approximation algorithm, identifying its tendency to diverge and fail on general inputs. It introduces three enhanced variants—coalition size constraints, multiple restarts, and adaptive scaling—demonstrating that scaling improves solution quality by a factor of 6.5 in the lower bound error, significantly outperforming the original algorithm.

ABSTRACT

The goal of this paper is to critically evaluate a heuristic algorithm for the Inverse Banzhaf Index problem by Laruelle and Widgrén. Few qualitative results are known about the approximation quality of the heuristics for this problem. The intuition behind the operation of this approximation algorithm is analysed and evaluated. We found that the algorithm can not handle general inputs well, and often fails to improve inputs. It is also shown to diverge after only tens of iterations. We present three alternative extensions of the algorithm that do not alter the complexity but can result in up to a factor 6.5 improvement in solution quality.

Motivation & Objective

• To critically assess the performance and limitations of the Laruelle-Widgrén heuristic for the inverse Banzhaf problem.
• To identify shortcomings in the original algorithm, such as divergence and poor handling of general inputs.
• To propose and evaluate three extensions that maintain computational complexity while improving solution quality.
• To quantify the improvement in approximation accuracy using error bounds and empirical testing.
• To explore the potential for combining improvements to achieve even better performance.

Proposed method

• Analyzes the heuristic mechanism of the Laruelle-Widgrén algorithm to identify failure modes, particularly around zero weights and convergence issues.
• Introduces three modifications: enforcing minimum coalition size, multiple restarts with random initialization, and adaptive scaling of weights via a factor.
• Uses a distance measure (L1 norm) to evaluate approximation quality against target power indices.
• Employs a controlled experiment with 5,000 samples of 8-player weighted voting games, tuned to optimal parameters (e.g., q=0.6, scaling factor=0.4).
• Compares all variants—including the base algorithm—on metrics including improvement rate, degradation rate, best performance frequency, and average error.
• Tracks the best solution across all runs to establish an upper bound on achievable performance.

Experimental results

Research questions

• RQ1 How does the original Laruelle-Widgrén algorithm perform across diverse input configurations, particularly in terms of convergence and solution quality?
• RQ2 To what extent do the proposed enhancements—coalition constraints, restarts, and scaling—improve the accuracy of the inverse Banzhaf approximation?
• RQ3 What is the quantitative impact of the scaling factor on the lower bound of approximation error compared to the original algorithm?
• RQ4 Can the improvements be combined effectively, and is there a synergistic effect in merging the three approaches?
• RQ5 Why does the original algorithm fail to improve in some cases, and how do the new variants mitigate this?

Key findings

• The original Laruelle-Widgrén algorithm often fails to improve inputs and can diverge after only tens of iterations, especially with high quotas.
• The scaling factor improvement reduces the average error to 0.0357 and the lower bound error to 0.0078, representing a 6.5-fold improvement over the base algorithm's lower bound.
• Scaling outperforms all other variants in terms of solution quality, achieving the best result in 80.0% of samples, compared to 3.6% for coalition and 3.0% for restart.
• The multiple restart variant never worsens the solution and performs at least as well as the base algorithm, though with less overall improvement.
• The minimum coalition size constraint performs best in 3.6% of cases and shows strong stability, particularly for low quotas.
• Combining all three improvements yields a further 20% reduction in the lower bound error, suggesting potential for a hybrid algorithm with superior performance.