Abstract
The Lioness Algorithm (LA) is a novel nature-inspired meta-heuristic optimization algorithm designed for finding optimal experimental designs. This package implements LA for continuous design spaces, optimal Latin hypercube designs (LHDs), and optimal order-of-addition (OofA) designs. Detailed methodologies and numerical simulations demonstrate its effectiveness in navigating complex design landscapes, outperforming or complementing traditional optimization methods.
Introduction
In fields ranging from engineering to pharmacology, the "design of experiments" (DOE) is a cornerstone of efficient research. The primary challenge lies in determining the optimal points for measurement to maximize information gain while minimizing costs. Finding the "optimal" design—one that fills a space evenly or maximizes the precision of a statistical model—is a notoriously difficult optimization problem. Traditional methods often struggle with high-dimensional spaces or become trapped in "local optima" (good solutions that aren't the best possible). The Lioness Algorithm (LA) addresses these challenges by mimicking the social and hunting behaviors of lions to explore and exploit rugged design landscapes effectively.
Methodology and Implementation
The 'LA' package leverages the Lioness Algorithm across three core experimental contexts, implemented using Rcpp and RcppArmadillo for high-performance computation while remaining accessible through a user-friendly R interface.
Latin Hypercube Designs (LHDs)
LHDs are essential for computer experiments to ensure uniform sampling across input factors. The LA_LHDC function optimizes these designs using various criteria, such as space-filling properties (maximin distance) or orthogonality (minimizing correlation between factors).
Order-of-Addition (OofA) Designs
In experiments where the sequence of components is critical, such as chemical mixing, OofA designs are used. The LA_OofAC function identifies D-optimal designs for these permutation-based problems, managing the rigid constraints where each level must appear exactly once.
Continuous Factor Designs
For standard regression models, the LA_OptC function finds optimal designs (D, A, or G-optimality) where factors can take any value within a defined range. It supports both first-order and second-order multiple linear regression models.
Key Insights and Results
The Lioness Algorithm balances exploration (searching new areas) and exploitation (refining known good solutions) through its "pride" structure. This versatility allows users to select from multiple optimality criteria, including phi_p for space-filling, AvgAbsCor for reducing factor correlation, and MaxPro for maximum projection properties.
Key results from the implementation include:
- Speed: Integration with C++ allows the algorithm to handle thousands of iterations (
N = 5000Lby default) relatively quickly. - Flexibility: The package supports a wide range of factor types and regression models.
- Ease of Use: Complex optimization tasks are reduced to single-line function calls, such as
LA_OptC(n=20, lb=c(1,0), ub=c(1,24)), which finds an optimal 20-run design within specified bounds.
Limitations and Future Work
As a meta-heuristic, the Lioness Algorithm is computationally more intensive than simple random sampling or basic exchange algorithms. Large designs may require significant CPU time. Additionally, the algorithm is stochastic; different runs might yield slightly different designs, meaning users may need to run the algorithm multiple times to ensure they have found a near-global optimum. Future directions could potentially expand to include mixed-level designs or designs for non-linear models.
Summary updated