Create a progressive adaptation schedule for use with sample_chain().

Returns a function (a schedule constructor) that, given the total number of warm-up iterations, produces a three-stage adaptation schedule:

Usage

progressive_adaptation_schedule(
  n_fixed_shape_iteration = 50L,
  n_diagonal_shape_iteration = 50L,
  scale_adapter_ = scale_adapter(),
  diagonal_shape_adapter_ = shape_adapter("variance"),
  dense_shape_adapter_ = shape_adapter("covariance")
)

Arguments

n_fixed_shape_iteration

Number of iterations in Stage 1 (scale only). Default 50.

n_diagonal_shape_iteration

Number of iterations in Stage 2 (scale + diagonal shape). Default 50.

scale_adapter_

Adapter object for the scale. Defaults to scale_adapter().

diagonal_shape_adapter_

Adapter object for the diagonal shape stage. Defaults to shape_adapter("variance").

dense_shape_adapter_

Adapter object for the dense shape stage. Defaults to shape_adapter("covariance").

Value

A function that accepts n_warm_up_iteration and returns a staged adapter list suitable for the adapters argument of sample_chain().

Details

• Stage 1 (n_fixed_shape_iteration iterations): adapt scale only, keeping the proposal shape fixed at the identity. This lets the step size stabilise before any covariance estimation begins.

• Stage 2 (n_diagonal_shape_iteration iterations): adapt scale and learn a diagonal proposal shape (per-dimension variances).

• Stage 3 (remaining iterations): adapt scale and learn a dense proposal shape (full covariance matrix).

If the sum of n_fixed_shape_iteration and n_diagonal_shape_iteration exceeds the total number of warm-up iterations, the two counts are reduced proportionally so that each gets half of the available iterations and Stage 3 is skipped.

Examples

target_distribution <- list(
  log_density = function(x) -sum(x^2) / 2,
  gradient_log_density = function(x) -x
)
withr::with_seed(876287L, {
  results <- sample_chain(
    target_distribution,
    initial_state = stats::rnorm(2),
    n_warm_up_iteration = 1000,
    n_main_iteration = 1000,
    adapters = progressive_adaptation_schedule()
  )
})