Introduction

Accelerated life testing (ALT) is used to obtain information about product lifetime under normal operating conditions by testing units under more severe stress conditions. Optimal experimental design can be used to determine the stress levels and allocation proportions that provide the most efficient estimation of a lifetime characteristic. This package provides methods to find locally C–optimal design and minimax design use particle swarm optimization (PSO). The locally optimal design is obtained when the model parameters are known, whereas the minimax design is used when the model parameters are unknown. In the minimax approach, the design is optimized under the worst-case parameter values, thus there are two layers of optimization. The inner layer to find parameter values uses Nelder–Mead algorithm, and the outer layer to find optimal design uses PSO. See (Lee, Chen, and Wong 2025) for more details.

This vignette demonstrates a complete workflow for generating and evaluating an optimal ALT design using this package.

The workflow consists of the following steps:

1. Specify the ALT design.
2. Set the particle swarm optimization (PSO) hyperparameters.
3. Specify initial values for the optimization.
4. Generate the optimal ALT design.
5. Check the optimality of the generated design.

1. Specify the ALT design

The first step is to determine the specifications of the design using .

design_info <- set_design_info(
    k_levels=2, 
    j_factor=1, 
    n_unit=300, 
    censor_time=183, 
    p=0.1, 
    use_cond=0, 
    sigma=0.6
)

print(design_info)
#> ALT design specifications
#> -----------------------------------------------
#> Optimality type: C-optimality
#> Number of stress levels: 2 
#> Number of factors: 1 
#> Number of testing units: 300 
#> Censoring time: 183 
#> Stress level at the use condition: (0) 
#> Stress range: [0, 1]
#> Lifetime percentile to be estimated at the use condition: 0.1 
#> Coefficients are reparameterized into failure probability? TRUE 
#> Scale parameter: 0.6

2. Set PSO hyperparameters

The optimal design is obtained using particle swarm optimization (PSO). The PSO algorithm requires several hyperparameters, which can be specified using .

pso_info <- pso_setting(
  n_swarm = 10,
  max_iter = 20,
  early_stopping = 10,
  tol = 0.0001,
  c1 = 2.05,
  c2 = 2.05,
  w0 = 1.2,
  w1 = 0.2,
  w_var = 0.8,
  vk = 4
)

print(pso_info)
#> PSO Hyperparameters
#> -----------------------------------------------
#> Number of particles: 10 
#> Maximum number of iterations: 20 
#> Frequency of checking optimality: 10 iterations
#> Convergence tolerance: 1e-04 
#> Cognitive acceleration coefficient c1: 2.05 
#> Social acceleration coefficient c2: 2.05 
#> Starting inertia weight w0: 1.2 
#> Ending inertia weight w1: 0.2 
#> Inertia weight decay fraction: 0.8 
#> Velocity clamping factor: 4

3. Specify initial values

In most applications, the default initialization is sufficient. User-specified initial values can be supplied when prior information about a suitable design or model parameters is available, using . The initial values to be specified includes particle positions for PSO and model parameters for multi–start Nelder–Mead in the case of minimax design. We create a swarm of 10 particles, and since the design has one factor with two stress levels, the particle has three elements, two of which are stress levels, and the other one is the weight of the first stress level.

init_swarm <- matrix(runif(10 * 3, min = 0, max = 1), nrow = 10, ncol = 3)
init_coef_mat <- rbind(
    c(1.01*10^-6, 0.999),
    c(9.99*10^-4, 0.999),
    c(1.01*10^-6, 0.701),
    c(9.99*10^-4, 0.701)
)

init_values <- initialize_values(
    init_swarm=init_swarm,
    init_coef_mat=init_coef_mat
)

print(init_values)
#> Initialized Values for PSO and Nelder-Mead Optimization
#> -----------------------------------------------
#> Number of initial particles for PSO: 10 
#> Particle size of PSO: 3 
#> Initial value (design) for Nelder-Mead optimization: () 
#> Multi-start Nelder-Mead optimization: 4 times

4. Generate the optimal ALT design

After specifying the design information, PSO settings, and initial values, an optimal ALT design can be generated using .

In this example, we consider a locally optimal design with a Log-normal failure-time distribution.

optimal_design <- find_optimal_alt(
    design_type="locally", 
    distribution="lognormal", 
    design_info=design_info, 
    pso_info=pso_info,
    coef = c(0.001, 0.9),
    highest_level = TRUE,
    verbose = FALSE
)

print(optimal_design)
#> PSO implementation
#> -----------------------------------------------
#> Iterations: 20 
#> Optimal particle: 1 0.4401631 0.2565753 
#> xive value: 0.01787859 
#> 
#> Optimality check
#> -----------------------------------------------
#> Number of model candidates: 1 
#> Max directional derivative: 1.014024

For demonstration purposes, the following sections assume that the optimization has been successfully completed and the resulting object is stored in .

5. Examine the optimization results

The resulting object is of class and contains the optimal design together with information from the optimization and optimality check.

The object can be inspected using:

summary(optimal_design)
#> Summary of generated optimal ALT design
#> -----------------------------------------------
#> X: Stress levels, W: Corresponding proportion
#>         [,1]      [,2]
#> X1 1.0000000 0.4401631
#> W  0.2565753 0.7434247
#> 
#> Objective Value: 0.01787859 
#> Max directional derivative: 1.014024

The summary displays the stress levels and their corresponding allocation proportions:

\[ X = \text{stress levels}, \qquad W = \text{corresponding allocation proportions}. \]

The objective value and maximum directional derivative are also reported. The optimality check can be visualized using the for objects.

plot(optimal_design)

Conclusion

The same workflow can be extended to other failure-time distributions, stress ranges, and optimization settings supported by the package.