Welcome to ARIS.jl!¶

GitHub Repository Paper and Reproducibility

Amortized Rare-Event Importance Sampling: one proposal, reused across failure thresholds.

ARIS.jl is a Julia package for estimating the probability that a system fails, when failure is too rare for naive Monte Carlo to see. It learns a joint model over system inputs and the robustness value they produce. Because the model is conditional on robustness, a single trained proposal yields an importance-sampling proposal for any threshold, by truncating the learned robustness distribution — rather than being specialised to the one threshold it was trained on.

Developed by the Stanford Intelligent Systems Laboratory (SISL).

Quickstart Installation

using ARIS
using Distributions, Random

include("examples/problems/toy.jl")

Random.seed!(1)
sys = UnimodalToy(3.0)

model = JointModel(Normal(), ConditionalEMGMM(2, 3, 3))

cv = ConditionalValidation(
    model = model,
    xdim = 2, depth = 1, sdim = 2,
    n_samples = 1000, n_iter = 25,
    sampling_strategy = QuantileSampling(0.2; target = 0.0),
)

train!(cv, sys)

is_estimate(cv.buffer; threshold = 0.0)

Robustness convention

A system returns a robustness value ρ for each input. Failure means ρ ≤ 0, and more negative values are deeper failures. Every estimator in the package targets ρ ≤ 0 unless you pass another threshold. See Rare-event estimation.

What amortization buys you¶

One fit, many thresholds

The proposal is a joint model q(x, r) over inputs and robustness. Restricting r ≤ γ turns it into a proposal aimed at the failure region {ρ ≤ γ}, for any γ, with no refitting and no extra simulator calls. See Amortized importance sampling.
An adaptive schedule, not a leap

Training walks the conditioning threshold down from the prior toward the target, so the proposal is never asked to jump straight into the rare set. See Methodology.

Where to go next¶

Quickstart: a complete run, end to end, with the numbers it produces.
Rare-event estimation: the problem, the robustness convention, and why importance sampling is needed.
Defining a system: the five methods your problem must implement.
Working with multiple thresholds: reading a whole curve off one run, and the assumptions that carries.
API reference: what the package exports, by signature.

Scope¶

Package API versus paper code

This site documents the package — what using ARIS gives you. The paper's multi-threshold estimator (pooled balance-MIS), its experiment drivers, metrics and figures live under paper/ and are not part of the public API. Paper and method says which is which.

The package is not registered in Julia's General registry; it is installed from a clone (Installation). Optional normalizing-flow and policy/MDP integrations exist behind environment-variable gates, are disabled by default, and are not used by any paper experiment.