[Paper Review] Engineering Art Galleries

Pedro J. de Rezende, Cid C. de Souza+4 more|arXiv (Cornell University)|Oct 31, 2014

Computational Geometry and Mesh Generation33 references4 citations

TL;DR

This paper presents a comprehensive experimental evaluation of algorithms for the Art Gallery Problem (AGP), comparing six implementations on 900 benchmark instances with up to 5,000 vertices. It identifies key algorithmic ingredients enabling state-of-the-art performance and introduces a new, optimized implementation that outperforms all prior approaches, significantly advancing the frontier of exact AGP solving in practice.

ABSTRACT

The Art Gallery Problem is one of the most well-known problems in Computational Geometry, with a rich history in the study of algorithms, complexity, and variants. Recently there has been a surge in experimental work on the problem. In this survey, we describe this work, show the chronology of developments, and compare current algorithms, including two unpublished versions, in an exhaustive experiment. Furthermore, we show what core algorithmic ingredients have led to recent successes.

Motivation & Objective

• To evaluate and compare existing implemented algorithms for the Art Gallery Problem (AGP) in a unified, reproducible experimental framework.
• To identify the core algorithmic components responsible for recent performance breakthroughs in solving AGP instances.
• To establish a benchmark suite of 900 problem instances with up to 5,000 vertices for future algorithmic comparison.
• To improve the leading implementations through joint optimization, resulting in a new state-of-the-art solver.

Proposed method

• Conducted a large-scale, unified experimental comparison of six implemented AGP algorithms on 900 benchmark instances with up to 5,000 vertices.
• Used a combination of integer linear programming (ILP) formulations and primal-dual methods, including the Braunschweig algorithm, to solve the AGP.
• Employed visibility polygon overlays and circular arc separation routines to accelerate visibility computation and constraint generation.
• Applied a simplex partitioning approach for continuous optimization to solve the fractional AGP, achieving fast convergence to near-optimal solutions.
• Integrated insights from multiple implementations to create a new, optimized solver that combines the best features of existing approaches.
• Evaluated algorithms under consistent conditions, including identical hardware, time limits, and statistical reporting, to ensure fair comparison.

Experimental results

Research questions

• RQ1 What are the key algorithmic components that have led to the recent dramatic improvements in solving AGP instances?
• RQ2 How do different algorithmic approaches—such as ILP, primal-dual, and continuous optimization—perform in practice on large-scale AGP instances?
• RQ3 What are the current performance limits of existing AGP solvers, and where do they fail?
• RQ4 Can the performance of leading implementations be further improved by combining their best features?
• RQ5 Are there inherent structural properties of certain polygons that make them intractable for current algorithms, even with small size?

Key findings

• The unified experimental evaluation revealed that the primal-dual Braunschweig algorithm and continuous optimization techniques based on simplex partitioning are among the most effective approaches for solving AGP.
• The new, jointly optimized implementation significantly outperforms all previous solvers, enabling the solution of instances with several thousand vertices in reasonable time on standard PCs.
• For polygons with up to 700 vertices, the continuous optimization approach achieved solutions with a duality gap of less than 0.01% and often found optimal or near-optimal solutions within 20 minutes.
• The study identified degenerate instances—such as the orthogonal polygon in Figure 8 with a unique optimal guard placement—that current algorithms cannot solve due to the lack of characterization for critical shadow alignment points.
• Despite significant progress, the existence of finite-time exact algorithms for AGP remains an open problem, particularly due to the potential need for irrational coordinates in optimal solutions.