Simulator and data-generation capabilities

ssm-simulators owns two related capabilities in the HSSM ecosystem: fast forward simulation from sequential-sampling models and generation of labeled datasets for likelihood-approximation networks. The first produces synthetic behavior; the second repeatedly uses simulation or an analytical estimator to construct training examples.

Direct simulation

The simulator layer covers classic diffusion models, collapsing-boundary and time-varying-drift variants, multi-choice accumulator models, attention models, and reinforcement-learning SSMs. Most users select a registered model and call the class-based Simulator interface. Model configurations describe parameters, bounds, choices, boundary/drift components, and parameter transforms; they do not contain a participant's concrete parameter values.

Use the basic tutorial for a first completed simulation. The configuration guide explains how to select and adapt registered models, while the custom-model guide covers runtime extensions.

Training-data generation

The data-generation layer samples valid parameter combinations, obtains likelihood labels through simulation-based KDE or compatible PyDDM solutions, and emits the feature/label arrays consumed by LANfactory. It also supports specialized targets such as choice- and omission-probability labels. LANfactory, not this package, owns network architecture and training behavior.

Use Generate training data for LANs for the task workflow. The data-generator API defines the programmatic surface.

Choosing a likelihood path

Simulation works across the broadest model set and can approximate a likelihood with KDE. PyDDM instead solves the Fokker–Planck equation for compatible single-particle, two-choice, Gaussian-noise models. The simulation and PyDDM explanation describes the trade-off and compatibility boundary; the KDE guide covers the simulation-based path.

Ecosystem boundary

ssm-simulators owns simulator definitions, model metadata, synthetic data, and training-data construction. LANfactory owns training and exporting neural likelihoods, and HSSM owns Bayesian inference. See the HSSM ecosystem map for the full artifact flow.