Overview of the package functionalities

[AlexanderFengler]

Hello maintainers,

thank you for providing the community with this simulator package. I am having some trouble orienting myself with all the different functionalities and model names.

Could you please give me a run-down of the basic functionality of this package and how it interfaces with the rest of the HSSM eco-system (LanFactory, LAN_pipeline_minimal)?

Best,
Alex

[AlexanderFengler]

🤖 HSSMeister (experimental): This answer was drafted by an AI assistant trained on the HSSM codebase and reviewed and approved by an HSSM maintainer before posting. We are testing this as a community support feature for the HSSM ecosystem — feedback on answer quality is very welcome!

---

ssm-simulators is a Python package for fast simulation of Sequential Sampling Models (SSMs). Its core function is to generate synthetic reaction time and choice data from a wide range of SSM variants — including the DDM, LBA, RACE models, and many others — via highly optimized simulators (many backed by C extensions for speed). The main entry point is the simulator() function, which takes a set of model parameters and returns simulated RT/choice arrays.

What it provides

• simulator() — the core function. Pass a model name, parameters, and number of trials; get back simulated (rt, response) data
• A large model library — 20+ SSM variants out of the box, from standard DDM to more exotic models like the angle DDM and full DDM
• Parameter support — handles both fixed and trial-wise (per-trial) parameter arrays
• High performance — designed for the kind of large-scale simulation needed to train likelihood approximation networks

Where it fits in the ecosystem

ssm-simulators is the simulation backbone for the broader set of packages maintained by the lab:

• Posterior predictive checks — after fitting an HSSM model, ssm-simulators is what powers hssm.plot_predictive() and related diagnostics. HSSM draws parameter samples from the posterior, passes them to ssm-simulators, and overlays the resulting simulated RT/choice distributions against your observed data. This is the primary way to assess whether your model actually captures the structure of your data
• LANfactory uses ssm-simulators to generate the large training datasets needed to train Likelihood Approximation Networks (LANs) — neural networks that learn to approximate the likelihood function of an SSM
• HSSM exposes ssm-simulators through hssm.simulate_data() for model recovery studies and power analyses, in addition to the posterior predictive workflow above
• LAN_pipeline_minimal handles the deployment side — loading trained LANs so HSSM can use them as fast likelihood approximators during inference

The typical flow: simulate with ssm-simulators → train a LAN with LANfactory → fit models with HSSM using those LANs via LAN_pipeline_minimal → evaluate fit with posterior predictives powered again by ssm-simulators.

Happy to go deeper on any specific part — parameter formats, available models, or how to plug simulated data into an HSSM workflow.

Best,
HSSMeister