ldmppr (version 1.1.2)

estimate_process_parameters: Estimate point process parameters using log-likelihood maximization

Description

Estimate spatio-temporal point process parameters by maximizing the (approximate) full log-likelihood using nloptr.

Usage

estimate_process_parameters(
  data,
  process = c("self_correcting"),
  grids,
  budgets,
  parameter_inits = NULL,
  delta = NULL,
  parallel = FALSE,
  num_cores = max(1L, parallel::detectCores() - 1L),
  set_future_plan = FALSE,
  strategy = c("local", "global_local", "multires_global_local"),
  global_algorithm = "NLOPT_GN_CRS2_LM",
  local_algorithm = "NLOPT_LN_BOBYQA",
  starts = list(global = 1L, local = 1L, jitter_sd = 0.35, seed = 1L),
  finite_bounds = NULL,
  refine_best_delta = TRUE,
  rescore_control = list(enabled = TRUE, top = 5L, objective_tol = 1e-06, param_tol =
    0.1, avoid_bound_solutions = TRUE, bound_eps = 1e-08),
  verbose = TRUE
)

Value

an object of class "ldmppr_fit" containing the best nloptr fit and (optionally) stored fits from global restarts and/or a delta search.

Arguments

data

a data.frame or matrix. Must contain either columns (time, x, y) or (x, y, size). If a matrix is provided without time, it must have column names c("x","y","size").

process

type of process used (currently supports "self_correcting").

grids

a ldmppr_grids object specifying the integration grid schedule (single-level or multi-resolution). The integration bounds are taken from grids$upper_bounds.

budgets

a ldmppr_budgets object controlling optimizer options for the global stage and local stages (first level vs refinement levels).

parameter_inits

(optional) numeric vector of length 8 giving initialization values for the model parameters. If NULL, defaults are derived from data and grids$upper_bounds.

delta

(optional) numeric scalar or vector. Used only when data does not contain time (i.e., data has (x,y,size)).

• If length(delta) == 1, fits the model once using power_law_mapping(size, delta).

• If length(delta) > 1, performs a delta-search by fitting the model for each candidate value and selecting the best objective. If refine_best_delta = TRUE and multiple grid levels are used, the best delta is refined on the remaining (finer) grid levels.

If data already contains time, delta is ignored when length(delta)==1 and is an error when length(delta)>1.

parallel

TRUE or FALSE specifying furrr/future to parallelize either: (a) over candidate delta values (when length(delta) > 1), and/or (b) over local multi-start initializations (when starts$local > 1), and/or (c) over global restarts (when starts$global > 1). For small problems, parallel overhead may outweigh speed gains.

num_cores

number of workers to use when parallel=TRUE and set_future_plan=TRUE.

set_future_plan

TRUE or FALSE. If TRUE and parallel=TRUE, set a temporary future plan internally and restore the previous plan on exit.

strategy

character string specifying the estimation strategy:

• "local": local optimization only (single-level or multi-level polish).

• "global_local": global optimization then local polish (single grid level).

• "multires_global_local": multi-resolution (coarsest uses global+local; refinements use local only).

global_algorithm, local_algorithm

NLopt algorithms to use for the global and local optimization stages, respectively.

starts

a list controlling restart and jitter behavior:

• global: integer, number of global restarts at the first/coarsest level (default 1).

• local: integer, number of local multi-starts per level (default 1).

• jitter_sd: numeric SD for jittering (default 0.35).

• seed: integer base seed (default 1).

finite_bounds

(optional) list with components lb and ub giving finite lower and upper bounds for all 8 parameters. If NULL, bounds are derived from parameter_inits. Global algorithms and select local algorithms in NLopt require finite bounds.

refine_best_delta

TRUE or FALSE. If TRUE and length(delta) > 1, performs refinement of the best delta across additional grid levels (if available).

rescore_control

controls candidate rescoring and bound-handling behavior in multi-resolution fitting. Can be either:

• a single logical value (toggle rescoring on/off while keeping defaults), or

• a named list with any of: enabled, top, objective_tol, param_tol, avoid_bound_solutions, bound_eps.

Defaults are: list(enabled = TRUE, top = 5L, objective_tol = 1e-6, param_tol = 0.10, avoid_bound_solutions = TRUE, bound_eps = 1e-8).

verbose

TRUE or FALSE indicating whether to show progress of model estimation.

Details

For the self-correcting process, arrival times must lie on \((0,1)\) and can be supplied directly in data as time, or constructed from size via the gentle-decay (power-law) mapping power_law_mapping using delta. When delta is a vector, the model is fit for each candidate value and the best objective is selected.

This function supports multi-resolution estimation via a ldmppr_grids schedule. If multiple grid levels are provided, the coarsest level may use a global optimizer followed by local refinement, and subsequent levels run local refinement only.

References

Møller, J., Ghorbani, M., & Rubak, E. (2016). Mechanistic spatio-temporal point process models for marked point processes, with a view to forest stand data. Biometrics, 72(3), 687-696. tools:::Rd_expr_doi("10.1111/biom.12466").

Examples

# Load example data
data(small_example_data)

# Define grids and budgets
ub <- c(1, 25, 25)
g  <- ldmppr_grids(upper_bounds = ub, levels = list(c(10,10,10)))
b  <- ldmppr_budgets(
  global_options = list(maxeval = 150),
  local_budget_first_level = list(maxeval = 50, xtol_rel = 1e-2),
  local_budget_refinement_levels = list(maxeval = 25, xtol_rel = 1e-2)
)

# Estimate parameters using a single delta value
fit <- estimate_process_parameters(
  data = small_example_data,
  grids = g,
  budgets = b,
  delta = 1,
  strategy = "global_local",
  global_algorithm = "NLOPT_GN_CRS2_LM",
  local_algorithm  = "NLOPT_LN_BOBYQA",
  starts = list(global = 2, local = 2, jitter_sd = 0.25, seed = 1),
  verbose = TRUE
)
coef(fit)
logLik(fit)

# \donttest{
# Estimate parameters using multiple delta values (delta search)
g2 <- ldmppr_grids(upper_bounds = ub, levels = list(c(8,8,8), c(12,12,12)))
fit_delta <- estimate_process_parameters(
  data = small_example_data, # x,y,size
  grids = g2,
  budgets = b,
  delta = c(0.35, 0.5, 0.65, 0.9, 1.0),
  parallel = TRUE,
  set_future_plan = TRUE,
  num_cores = 2,
  strategy = "multires_global_local",
  starts = list(local = 1),
  refine_best_delta = FALSE,
  verbose = FALSE
)
plot(fit_delta)
# }