tags:
- lan
- ssm
- hssm
- angle
- collapsing-bound
library_name: onnx
license: bsd-2-clause
angle_extended (LAN)
Likelihood approximation network for the angle_extended sequential sampling model: the standard angle model -- constant drift v with a linearly collapsing decision bound whose collapse angle is theta -- with the drift bounds widened from (-3, 3) to (-6, 6) for designs that produce strong evidence. Parameters are v, a, z, t, theta, in that order, and the network was trained on the full ssm-simulators bounds for the model. Recovery verdict: PASS. Judged alone on a 240-fit sweep (12 design cells x 20 paired datasets, 81 family-corrected tests) this network records zero coverage failures, zero bias failures, and zero cells the design could not identify, with 84% of parameter records converged (r-hat <= 1.01, bulk ESS >= 400). The extended drift range is exercised, not merely declared: density checks and recovery truths reach |v| near 4.5 and beyond. The sweep contains no analytical or blackbox reference arm -- none exists for a collapsing- bound model -- so, as with every LAN in this family, coverage is judged against nominal rather than against an exact-likelihood ceiling. Density: the worst Hellinger distance to a 20k-trial simulation is 0.054, which is BELOW the Monte Carlo sampling floor at that sample size (ratio 0.92 against a gate bar of 3.0) -- the network is closer to the simulator than the simulator is to itself at 20k trials. Total probability mass stays within 0.4% of 1.0 on every draw. A minority of truth draws produce genuinely hard posteriors -- long sampler trajectories, occasional non-convergence; wall time and convergence correlate at -0.50 across the sweep. This is posterior geometry at specific truths, not a defect localised to the network, and no design cell fell below the judgement threshold because of it.
Architecture
- Network Type: lan
- Layer Sizes: [128, 128, 128, 128, 1]
- Activations: ['tanh', 'tanh', 'tanh', 'tanh', 'linear']
Training
- Epochs: 20
- Optimizer: adam
- Learning Rate: 0.001
- Loss: huber
Usage with HSSM
import hssm
# p_outlier=0.0 is deliberate. HSSM attaches a 5% lapse by default, but the
# training data contains no lapses, so the default misspecifies every fit --
# the bias stays fixed while the posterior sd shrinks as 1/sqrt(n), so
# coverage gets WORSE as the dataset grows. Set a lapse only if your data
# actually has one.
model = hssm.HSSM(data=data, model="angle_extended", p_outlier=0.0)
# On an HSSM release that predates the built-in config, pass the network
# and the box explicitly. Keep the parameter order exactly as written: it
# is the ONNX input column order, and a wrong order does not raise -- it
# silently evaluates the wrong likelihood.
model = hssm.HSSM(
data=data,
model="angle_extended",
loglik="angle_extended.onnx",
loglik_kind="approx_differentiable",
p_outlier=0.0,
model_config={
"list_params": ["v", "a", "z", "t", "theta"],
"backend": "jax",
"bounds": {
"v": (-6.0, 6.0), "a": (0.3, 3.0), "z": (0.1, 0.9),
"t": (0.001, 2.0), "theta": (-0.1, 1.3),
},
},
)
Citation
If you use this model, please cite: