Comparison to jaxsr¶
jaxsr ("a symbolic regression engine for
JAX") is a broad, mature general-purpose symbolic-regression toolkit;
jax-ibart is a faithful, specialized implementation of one method
(iBART/OIS). They overlap substantially but their distinctive strengths differ.
Shared (real overlap)¶
Both are JAX (JIT + GPU), use LASSO sparse selection, provide conformal and bootstrap uncertainty quantification and OLS coefficient intervals, score with BIC, export to SymPy, draw parity plots, expose a scikit-learn-style API, and can do explicit SINDy-style ODE discovery.
Only jax-ibart¶
| Capability | Why it matters |
|---|---|
| Iterative operator-composition library | Grows the library by composing operators and pruning between rounds, reaching deep descriptors like (exp(x1)-exp(x2))^2 without enumerating the O(10^17) candidate space. jaxsr uses a fixed, user-specified basis. |
| Nonparametric BART-G.SE screening (PAN) | Bayesian-tree, monotone-invariant screening robust to the heavy collinearity that iterated operators create; this is what makes the iterative growth tractable. |
| Dimensional-analysis (unit) filter | Drops dimensionally invalid descriptors (volume + speed). |
| Differentiable selection backend | Hard-concrete L0 gates: selection that gradients flow through. |
| Implicit / null-space discovery | Finds algebraic constraints g(x)~0 (the DAE foundation; recovers Michaelis-Menten exactly from noiseless data). |
| Selection-stability + BART posterior UQ | Per-descriptor bootstrap inclusion frequency; Bayesian-tree posterior predictive intervals. |
| Validated against reference R-iBART | Oracle cross-check; Liu 2022 / Ye 2024 (see References). |
Only jaxsr¶
| Capability | Notes |
|---|---|
| Symbolic classification | binary/multiclass. jax-ibart is regression-only. |
| Physical constraints | monotonicity, convexity/concavity, bounds, sign, linear. jax-ibart has none (only the unit filter). |
| Bayesian Model Averaging / Pareto ensembles | average over models; complexity-accuracy front. jax-ibart selects one sparse model. |
| Adaptive sampling / DOE | uncertainty / leverage / gradient acquisition functions. |
| Broader UQ + criteria | jackknife+ conformal, mean-response confidence bands, AIC/AICc. |
| Tooling & maturity | LaTeX export, JSON persistence, full sklearn (set_params / GridSearchCV / Pipeline), CLI tools, and a rich visualization suite (Pareto, residual, coefficient paths, forest plots, BMA weights). |
Which to use¶
jaxsr-- a flexible, full-featured general SR toolkit: when you can specify the basis and want constraints, Bayesian model averaging, design of experiments / active learning, classification, rich UQ, or visualization.jax-ibart-- when you specifically want the iBART method: adaptive library growth for hard high-order nonlinear descriptors, BART-based screening, dimensional-analysis-aware discovery, a differentiable selection backend, fidelity to the published catalysis results, or the implicit-constraint / DAE direction.
Capabilities for
jaxsrare summarized from its README; see the jaxsr repository for the authoritative, current feature set.