Set up

1. Install uv https://docs.astral.sh/uv/getting-started/installation/ if you don't have one.
2. uv sync to install dependencies

Demonstration

See demonstration.ipynb

Running experiments

Run experiment with our method

uv run experiments/train.py --simulator lv --method dro_ub --epsilon 1.0 --n 1024

• simulator: lv stands for "Lotka–Volterra". You can also choose tm (two moons), slcp (SLCP), and ik (Inverse Kinematics).
• method: dro_ub is our method.
• epsilon: Hyperparameters for our method. Larger epsilon corresponds to larger Wasserstein ball, leading to more conservative inference.
• n: simulation budget (number of $(\theta, x)$ pair to sample for training)
• Other parameters:
  • seed: setting seed for experiments (sampling from simulator, weight initialisation, batch selection...)
  • epochs: number of epochs to train.
  • val_rate: rate of validation data.

Run experiment with our method + epsilon selection.

uv run experiments/train_and_optimise_hyperparams.py --simulator lv --n 1024 --metric kl_cal --optimiser BO --eps_min 0.001 --eps_max 10 --maxiter 10

• metric: metric used to pick epsilon. Our propose $kl_\mathrm{cal}$ (kl_cal) is used as a default. You can use other metric such as NLPD.
  • We always use 90% of the data to train and 10% of the data to evaluate the metric.
  • After the optimal epsilon is identified, we train using all the data with the optimal epsilon.
• optimiser: For efficient search of optimal epsilon, we use Bayeisan optimisation.
• eps_min, eps_max: Lower and upper bound for epsilon to be searched.
• maxiter: Number of iteration for Bayesian optimisation.