Abstract
The 'LHD' package contains different algorithms and construction methods for optimal Latin hypercube designs (LHDs) with flexible sizes. It is capable of generating maximin distance LHDs, maximum projection LHDs, and orthogonal and nearly orthogonal LHDs. Detailed comparisons and summary of all the algorithms and construction methods in this package can be found at Hongzhi Wang, Qian Xiao and Abhyuday Mandal (2021) doi:10.48550/arXiv.2010.09154. This package is particularly useful in the area of Design and Analysis of Experiments (DAE), specifically for the design of computer experiments.
Introduction: The Challenge of Computer Experiments
In fields such as engineering, climate modeling, and drug discovery, researchers frequently rely on complex computer simulations. These simulations are often computationally "expensive," making it impossible to test every possible combination of input variables. To extract the maximum information from a limited number of runs, researchers employ Design of Experiments (DOE).
A simple random sample can lead to undesirable clustering or large gaps in the input space. A Latin Hypercube Design (LHD) addresses this by ensuring that for each individual variable, every "level" is sampled exactly once. However, standard LHDs may still suffer from high factor correlation or poor space-filling properties. The ‘LHD’ package provides a robust suite of algorithms to construct "optimal" designs that are either perfectly spread out (maximin distance) or statistically independent (orthogonal).
Methods and Construction
The package implements a diverse array of construction methods and optimization heuristics to evolve random designs into optimal ones:
- Heuristic Search Algorithms: Includes Simulated Annealing (SA), Genetic Algorithms (GA), and Particle Swarm Optimization (LaPSO) to minimize specific criteria.
- Algebraic Construction: Provides methods based on Orthogonal Arrays (OA) and Good Lattice Points (GLP) for designs with specific mathematical properties like perfect orthogonality.
- Specialized Transformations: Implements the Williams Transformation and various methods for Sliced LHDs, useful for experiments run in multiple batches.
Optimality Criteria
Beyond construction, the package evaluates designs using rigorous metrics. One fundamental metric is the Maximin Distance, which ensures points are as far apart as possible to fill the space efficiently. This often requires calculating inter-site distances between rows of the design matrix.
Other criteria include Maximum Projection (MaxPro), which maintains good properties even if some variables are found to be unimportant, and Average Absolute Correlation, which minimizes the statistical dependence between factors.
Key Insights: Efficiency and Flexibility
A major contribution of the ‘LHD’ package is its centralization of disparate research into a single, user-friendly library. It offers significant flexibility in design size; unlike many traditional methods that require specific run counts (e.g., prime numbers), ‘LHD’ allows engineers to tailor designs to their specific budget.
One of the most significant features is the "Fast" approach to design generation. The FastMmLHD function allows for the creation of high-quality designs in high-dimensional spaces where traditional search algorithms might stall.
Results and Practical Considerations
The package successfully implements over 20 distinct functions for design generation and evaluation. By providing standardized functions like phi_p (for space-filling) and AvgAbsCor (for orthogonality), it allows researchers to quantitatively compare different strategies.
However, users must navigate certain trade-offs. Heuristic algorithms like Genetic Algorithms or Particle Swarm Optimization can be CPU-intensive for large designs. Additionally, some specialized orthogonal construction methods still carry inherent mathematical constraints on the number of runs. Despite these limitations, ‘LHD’ remains a comprehensive "one-stop" utility for the modern design of computer experiments.
Summary updated