Create and Integrate Your Custom Model

Introduction

The FishSET R package provides a robust framework for discrete choice modeling (DCM), utilizing RTMB (R Template Model Builder) for automatic differentiation and optimized parameter estimation. While functions like fishset_design() and fishset_fit() offer a solid foundation for conditional logit and expected profit models, we recognize that specialized research often requires custom model structures.

To maximize computational performance and ensure compatibility with downstream FishSET tools (like our policy function and database loggers), custom models must be integrated directly into the RTMB framework. We encourage users to develop their own objective functions and submit them to be merged into the official FishSET toolkit.

This vignette outlines the workflow for developing a custom model by leveraging existing FishSET data pipelines and writing an RTMB negative log-likelihood function. Reach out to our team at nmfs.fishset@noaa.gov with any questions.

Development guidelines

Fork the FishSET repo and clone the forked repository to your computer. More information on “forking” repositories and the GitHub workflow can be found here.
Open the FishSET_RPackage.Rproj file.
Develop your custom data processing and RTMB objective function following the steps below.
Modify fishset_fit.R and fishset_design.R locally to include your model.
Review and test code on local datasets.
Commit, push changes, and submit a pull request.

Step 1: Extract Formatted Data

Rather than formatting data manually, rely on the built-in format_model_data() function to aggregate your main dataset, distance matrices, expectations, and gridded data.

This function saves the combined dataset as a .qs2 or .rds file in the Models/FormattedData directory inside your project folder. Load this dataset into your environment to begin designing your model matrix.

# Example: Loading formatted data
project_dir <- file.path(locproject(), "MyProject")
formatted_dir <- file.path(project_dir, "Models", "FormattedData")
my_data <- qs2::qs_read(file.path(formatted_dir, "MyProjectLongFormatData.qs2"))

Step 2: Mock the Design Object

The fishset_fit() function expects a specific list structure, of class fishset_design, to perform estimations. If your model requires predictors or matrix shapes that the standard fishset_design() function cannot currently generate, write a custom data processing script that builds a list mirroring this required structure.

Your design object must contain the following elements:

y: A numeric vector indicating the binary choice variable (0 or 1).
X: A sparse model matrix containing your formatted predictors.
settings: A list containing dimensions such as N_obs (number of observations), J_alts (number of alternatives), and K_vars (number of parameters).
ids: A list containing unique observation IDs and zone identifiers to be used for post-estimation predictions.

Step 3: Write the RTMB Objective Function

The core of FishSET’s speed relies on RTMB::MakeADFun and stats::nlminb. To integrate your model, you must write a negative log-likelihood (NLL) function compatible with this framework.

Inside fishset_fit(), parameter estimation relies on two main components passed to RTMB:

data_list: A list containing all known data (X matrix, observations, dimensions).
start_pars: A list of initial parameter values to be optimized (e.g., betas).

Your custom function must extract these elements using RTMB::getAll() and calculate the NLL.

Template: Custom Objective Function

# Example custom model template for an RTMB NLL function
nll_func_custom <- function(pars) {
  # Unpack data_list and parameters into the function environment
  RTMB::getAll(data_list, pars)
  
  # Perform matrix multiplication (e.g., sparse design matrix %*% parameters)
  v <- X %*% betas
  
  # Extract utility of chosen alternatives
  v_chosen <- v[chosen_lin_idx]
  
  # Reshape utility vector for log-sum-exp calculations
  dim(v) <- c(J_alts, N_obs)
  
  # Calculate log-sum-exp (consider robust numeric scaling if necessary)
  log_sum_exp <- log(RTMB::colSums(exp(v)))
  
  # Calculate the negative log-likelihood
  nll <- -sum(v_chosen - log_sum_exp)
  
  return(nll)
}

Important Note: When writing an RTMB function, you must use RTMB compatible math functions (e.g., RTMB::dnorm, RTMB::rowSums) rather than standard base R variants to ensure automatic differentiation works properly.

Step 4: Integrate and Submit

Once you have a working design object and an RTMB objective function, you can integrate your code into the package.

Modify fishset_fit.R: Add a new if/else block within the fishset_fit() function to handle your custom model type. Set up your unique data_list and start_pars, and define your custom nll_func within this block.
Standardize Outputs: Ensure your code allows RTMB::MakeADFun and stats::nlminb to run, producing a result formatted as class fishset_fit. This guarantees your model outputs a coefficient table, diagnostics, AIC/BIC, and predicted probabilities that seamlessly integrate with the project database.
Submit a Pull Request: Give the Pull Request an informative name and provide a brief description of the mathematical structure of your model. Our team will review the code and notify you of any requested changes prior to merging.

FishSET Development Team

2026-09-04