Too Many Alternatives to Handle

This repository contains two approaches for Bayesian estimation of large discrete-choice models: an Adaptive Sampler based on sampled choice sets, and TurboLFI/BOLFI, a likelihood-free inference approach. Each method has its own environment and run instructions below.

Authors and Supervision

• Anne-Valérie Preto — anne-valerie.preto@epfl.ch
• Xinwei Li — xinwei.li@u.nus.edu

Supervised by Prof. Michel Bierlaire and Prof. Prateek Bansal.

1. Adaptive Sampler

This repository contains an Adaptive Sampler implementation for discrete choice models using PyMC and a custom adaptive alternative sampling step.

1.1. Environment Setup

To ensure reproducibility (matching the Scitas computation environment), create a Conda environment using the provided requirements:

# 1. Create the environment
conda create -n adasampler python=3.11 -y
conda activate adasampler

# 2. From the repository root, install dependencies
pip install -r requirements_repro.txt

1.2. Data Configuration

The sampler expects CSV files following the naming scheme: choices_J{J}_P{P}_N{N/1000}k_I{I}C{C}M{M}.csv (for example, choices_J100_P10_N10k_I0C0M0.csv).

• Bundled path: data/data/final_2310/
• Change path: Pass data_dir when calling load_data() from Python, or add a directory to candidate_dirs in adaptive_sampler/data.py.

1.3. Running the Model

Run the sampler from the adaptive_sampler/ directory (this also keeps the log and default output/ directory together):

python run_model.py \
  --sample_size 2 \
  --L 5 \
  --m 5 \
  --n_steps 10000 \
  --burnin 0 \
  --tune 1000 \
  --seed 42 \
  --J 100 \
  --N 10000 \
  --P 10 \
  --I 0 \
  --C 0 \
  --M 0 \
  --outdir output \
  --scaling scale \
  --rescale_trace \
  --output_suffix "_260603"

Parameter Reference

Parameter	Meaning
sample_size	Number of alternatives in the sampled set (must be 2 for this protocol).
L	MTM Parameter: Number of proposals for the first stage.
m	MTM Parameter: Number of proposals for the second stage.
n_steps	Number of MCMC draws to keep.
burnin	Number of initial draws to discard (metadata only).
tune	Number of tuning/warmup steps for PyMC.
seed	Random seed for reproducibility.
J, N, P	Data config: Alternatives (100-500), Observations (10k-50k), Parameters (10-40).
I, C, M	Data config: Imbalance, Correlation, Randomness (0 or 1).
outdir	Directory where .nc traces and .json reports are saved.
scaling	Scaling mode (scale or normalize) to improve convergence.
rescale_trace	Automatically convert parameters back to original units before saving.
output_suffix	String appended to output filenames for versioning.

1.4. Outputs

Results are saved in the defined --outdir:

• trace_...nc: The posterior samples.
• reproducibility_info_...json: System info and validation logs.
• scale_params_...json: Parameters used for feature scaling.

2. TurboLFI / BOLFI

This directory contains the CPU implementation of TurboLFI/BOLFI for the synthetic multinomial-logit experiments. The entry point is run_bolfi_canonical_v2.py; there is no run_bolfi_canonical.py in this repository.

2.1 Environment

TurboLFI has its own Python 3.11 environment. Create it from the repository root so that it stays separate from the adaptive-sampler environment:

conda create -y -n turbolfi python=3.11
conda activate turbolfi
python -m pip install --upgrade pip
python -m pip install -r TUBOLFI/requirements.txt

Use requirements.txt on macOS and Linux. requirements-lock.txt is a full package snapshot from the Windows/Intel machine used for the recorded results; it includes platform-specific packages such as intel-cmplr-lib-ur, which are not published for macOS and therefore cannot be installed there. The shorter requirements.txt contains the portable direct dependencies required by the runner.

Verify the installation before submitting a long run:

cd TUBOLFI
python -c "import torch, botorch, gpytorch, pymc, arviz; print(torch.__version__)"
python run_bolfi_canonical_v2.py --help

2.2 Data and outputs

For --J 100 --N 10000 --P 10, the runner reads these bundled files:

data/restaurants_J100_P10_N10k_I0C0M0.csv
data/choices_J100_P10_N10k_I0C0M0.csv

It writes results directly below:

final_result/SA/J100_P10_N10k_I0C0M0_D5+6_SIM30/

The default simulator is SA. A full run trains the GP and then runs two four-chain PyMC MCMC passes (early_stop and iter_max when convergence is fixed). It overwrites checkpoints and MCMC artifacts for the selected configuration. Use a separate copy of final_result/ if its checked-in results must be retained.

2.3 Run on Slurm

After activating the turbolfi environment and changing to TUBOLFI/, the requested command is:

srun /usr/bin/time -v python run_bolfi_canonical_v2.py --J 100 --N 10000 --P 10

For an explicitly reproducible training seed, use:

srun /usr/bin/time -v python run_bolfi_canonical_v2.py --J 100 --N 10000 --P 10 --seed 42

script_Bolfi.run is the corresponding batch-script template. Before sbatch, set its #SBATCH --chdir, virtual-environment activation path, and #SBATCH --output path to locations on the target cluster. The file now calls the existing _v2 runner and only uses supported command-line flags.

Useful options are --simulator_type {SA,all}, --seed, and --convergence_type {fixed,auto}. The full fixed run is computationally substantial: its BO budget is 150 * P iterations and it subsequently performs two MCMC jobs with 4 chains, 2,500 tuning draws, and 5,000 retained draws each. See reproducible_setting.md for the canonical settings and retained-result summary.