Logistic-Gated Operators in Symbolic Regression

• Logistic-Gated Operators (LGO) are differentiable logistic gates integrated into symbolic regression to learn thresholds and steepness, providing auditable, unit-aware cut-points.
• LGO employs a strongly-typed genetic programming framework that standardizes inputs, learns thresholds in z-space, and inverts them to natural units for clinical audits.
• LGO optimizes parameters through a hybrid evolutionary and local-search approach, resulting in sparse regime-switching models that balance interpretability and predictive accuracy.

Logistic-Gated Operators (LGO) are differentiable gates with learnable location and steepness embedded as typed primitives in symbolic regression and mapped back to physical units for audit. They were introduced to address a specific limitation of symbolic regression: readable equations often struggle to encode unit-aware thresholds and conditional logic. In the reported formulation, thresholds are learned in standardized space, inverted to natural units, and compared to curated clinical anchors, yielding compact symbolic equations with explicit, unit-aware thresholds that can be audited against clinical anchors (Deng et al., 5 Oct 2025).

1. Primitive definition and limiting behavior

At the core of LGO is a differentiable logistic gate

$g(x;\alpha,\beta)=\frac{1}{1+e^{-\beta(x-\alpha)}}$

where $\alpha\in\mathbb{R}$ is the location (threshold) in standardized (z) space and $\beta>0$ is the steepness (inverse “softness”).

Two canonical 1-input forms are defined. The soft-gate is

$$\mathrm{LGO}_{\text{soft}}(x;\alpha,\beta)=x\cdot g(x;\alpha,\beta),$$

and the hard-gate is

$$\mathrm{LGO}_{\text{hard}}(x;\alpha,\beta)=g(x;\alpha,\beta).$$

In the $\beta\to\infty$ limit,

$g(x;\alpha,\beta)\to \mathbf{1}\{x>\alpha\},$

and

$$\mathrm{LGO}_{\text{soft}}\to x\cdot \mathbf{1}\{x>\alpha\}.$$

This gives LGO a continuous-to-discrete transition: finite $\beta$ yields differentiable switching, while large $\beta$ approaches a Heaviside-style threshold. In the symbolic-regression setting, this is used to represent regime switching without abandoning gradient-compatible parameter refinement (Deng et al., 5 Oct 2025).

2. Typed embedding and unit-aware thresholds

LGO is implemented inside a strongly-typed genetic-programming (STGP) engine with three disjoint types:

Type	Role
Feat	raw or subexpression values
Pos	positive steepness parameters ($\alpha\in\mathbb{R}$0, reparameterized via softplus)
Thr	threshold parameters ($\alpha\in\mathbb{R}$1, constrained to $\alpha\in\mathbb{R}$2 in z-space)

All feature inputs $\alpha\in\mathbb{R}$3 are z-score standardized at training time:

$\alpha\in\mathbb{R}$4

Gates $\alpha\in\mathbb{R}$5 operate in this standardized space. After structure search and $\alpha\in\mathbb{R}$6 fitting, thresholds $\alpha\in\mathbb{R}$7 in z-space are inverted to natural units for audit:

$\alpha\in\mathbb{R}$8

This design makes threshold learning compatible with typed program synthesis while preserving a direct mapping to domain units. The reported use case is clinical audit: a threshold learned in standardized space can be rendered as a cut-point in mmHg, mmol/L, mg/dL, or cm, rather than remaining an opaque latent parameter (Deng et al., 5 Oct 2025).

3. Objective, search, and parameter learning

The overall objective is a mean-squared-error (MSE) on the training fold, with a complexity (node-count) penalty enforced via Pareto selection:

$\alpha\in\mathbb{R}$9

The optimization pipeline combines discrete structure search with continuous parameter adjustment. Search is performed by tournament-based GP with subtree crossover/mutation. Micro-mutations randomly perturb $\beta>0$0 at each LGO node during evolution. Local refinement is then applied after structure selection by coordinate descent on $\beta>0$1 while holding the tree fixed (Deng et al., 5 Oct 2025).

For a 1-input gate $\beta>0$2, the reported gradients are

$\beta>0$3

For $\beta>0$4, an extra factor of $\beta>0$5 multiplies these derivatives. The practical consequence is that LGO parameters are not treated as static constants inside the tree; they are optimized jointly with the symbolic structure by a hybrid evolutionary and local-search procedure.

4. Soft versus hard variants and the pruning mechanism

The two variants induce different structural biases. Soft-gates multiply the entire subexpression by a sigmoid mask. They tend to produce many “partial” switches and thus more gates per model. Hard-gates output only the sigmoid mask. They encourage sparse, near-binary switching and, in practice, prune unused gates automatically when regime switching is not supported by data (Deng et al., 5 Oct 2025).

Among the top-100 models, the reported gate counts are:

Dataset	Soft-gates median	Hard-gates median
ICU	10.0 gates/model	4.0 gates/model
NHANES	12.5 gates/model	5.0 gates/model

On smooth tasks, the same sparsity pressure becomes more pronounced. On datasets with globally smooth relations, such as the Hydraulic system, $\beta>0$6 drops to approximately 0 gates and reverts to the smooth base set. This directly addresses a common misconception that introducing gates necessarily forces threshold structure into every discovered equation. In the reported experiments, thresholds appear only when supported by regime shifts in the data; otherwise the gates are pruned and the model remains algebraic (Deng et al., 5 Oct 2025).

5. Clinical audit, anchor agreement, and empirical results

A central feature of LGO is that learned $\beta>0$7 thresholds are converted to raw units by

$\beta>0$8

and then compared against curated clinical anchors, including MAP 65 mmHg, lactate 2 mmol/L, SBP 130 mmHg, and fasting-glucose 100 mg/dL. Agreement is reported as relative deviation

$\beta>0$9

and traffic-lighted as green for $$\mathrm{LGO}_{\text{soft}}(x;\alpha,\beta)=x\cdot g(x;\alpha,\beta),$$0, yellow for $$\mathrm{LGO}_{\text{soft}}(x;\alpha,\beta)=x\cdot g(x;\alpha,\beta),$$1, and red for $$\mathrm{LGO}_{\text{soft}}(x;\alpha,\beta)=x\cdot g(x;\alpha,\beta),$$2 (Deng et al., 5 Oct 2025).

On ICU and NHANES, among assessable gates, approximately 71% fall within 10% and 100% within 20% of guideline anchors. In the abstract this is stated as 71% (5/7) of assessed thresholds within 10% of guideline anchors and 100% within 20%. The reported interpretation is that LGO can recover clinically plausible cut-points while keeping the thresholds explicit and auditable.

The primary regression results are as follows:

Dataset	$$\mathrm{LGO}_{\text{soft}}(x;\alpha,\beta)=x\cdot g(x;\alpha,\beta),$$3	$$\mathrm{LGO}_{\text{soft}}(x;\alpha,\beta)=x\cdot g(x;\alpha,\beta),$$4
ICU	$$\mathrm{LGO}_{\text{soft}}(x;\alpha,\beta)=x\cdot g(x;\alpha,\beta),$$5, RMSE $$\mathrm{LGO}_{\text{soft}}(x;\alpha,\beta)=x\cdot g(x;\alpha,\beta),$$6, MAE $$\mathrm{LGO}_{\text{soft}}(x;\alpha,\beta)=x\cdot g(x;\alpha,\beta),$$7, median 4.0 gates	$$\mathrm{LGO}_{\text{soft}}(x;\alpha,\beta)=x\cdot g(x;\alpha,\beta),$$8, median 10.0 gates
NHANES	$$\mathrm{LGO}_{\text{soft}}(x;\alpha,\beta)=x\cdot g(x;\alpha,\beta),$$9, RMSE $$\mathrm{LGO}_{\text{hard}}(x;\alpha,\beta)=g(x;\alpha,\beta).$$0, MAE $$\mathrm{LGO}_{\text{hard}}(x;\alpha,\beta)=g(x;\alpha,\beta).$$1, median 5.0 gates	$$\mathrm{LGO}_{\text{hard}}(x;\alpha,\beta)=g(x;\alpha,\beta).$$2, median 12.5 gates

For ICU, the target is described as composite risk; regression. For NHANES, it is metabolic score; regression. Baselines including PySR, Operon, and RILS-ROLS lie in a similar accuracy envelope but with no auditable gates (Deng et al., 5 Oct 2025).

The paper provides an example top-1 $$\mathrm{LGO}_{\text{hard}}(x;\alpha,\beta)=g(x;\alpha,\beta).$$3 equation for NHANES:

$$\mathrm{LGO}_{\text{hard}}(x;\alpha,\beta)=g(x;\alpha,\beta).$$4

The associated thresholds are reported as TG approximately 200 mg/dL, SBP approximately 128 mmHg, HDL approximately 40 mg/dL, and waist approximately 94 cm, each within 1–2% of clinical anchors. This example makes the intended representational split explicit: discrete regime indicators are handled by gates, while continuously varying effects remain in a smooth subexpression.

6. Interpretability as a modeling constraint and governance-ready deployment

The reported claim is not merely that LGO yields post-hoc explanations. The abstract states that it turns interpretability “from a post-hoc explanation into a modeling constraint.” In this formulation, thresholds are first-class symbolic objects with typed semantics, learned parameters, and an explicit inverse mapping to native units (Deng et al., 5 Oct 2025).

Because thresholds are explicit, unit-aware, and compared to domain guidelines, LGO models are described as governance-ready. The stated reasons are concrete: clinicians and regulators can directly audit cut-points in their native units, recalibrate anchors as guidelines evolve, and trace how each gate enters the final formula. A plausible implication is that the audit trail is tied to the symbolic expression itself rather than to a separate explanation layer. This differs from many interpretable-model claims in which threshold semantics emerge only after fitting.

The same section of the source also clarifies the role of parsimony. Selective gating yields parsimonious formulas: thresholds appear only when supported by regime shifts in the data. On predominantly smooth tasks, gates are pruned, preserving parsimony. This places LGO between two familiar extremes in symbolic regression: purely smooth algebraic expressions and brittle hard-threshold rule systems. The hard-gate variant, in particular, is presented as a practical calculus for regime switching with explicit auditability.

7. Related uses of logistic gating in adjacent literatures

The phrase “logistic-gated” appears in other research areas, but the objects and objectives differ materially from LGO in symbolic regression. In multinomial logistic mixture-of-experts, a modified sigmoid gate is defined by

$$\mathrm{LGO}_{\text{hard}}(x;\alpha,\beta)=g(x;\alpha,\beta).$$5

with latent scores $$\mathrm{LGO}_{\text{hard}}(x;\alpha,\beta)=g(x;\alpha,\beta).$$6. That work studies identifiability, convergence, temperature effects, and sample-complexity rates for routing among experts, including a Euclidean-score alternative that restores polynomial rates under temperature scaling (Pham et al., 1 Feb 2026). The shared element is logistic gating; the difference is that the gate there is a normalized expert-weighting mechanism rather than a typed symbolic primitive mapped back to physical units.

A separate usage arises in state-inclusive Koopman approximations, where multivariate logistic observables are constructed as products of one-dimensional logistic functions,

$$\mathrm{LGO}_{\text{hard}}(x;\alpha,\beta)=g(x;\alpha,\beta).$$7

and included in a lifted state

$$\mathrm{LGO}_{\text{hard}}(x;\alpha,\beta)=g(x;\alpha,\beta).$$8

There the emphasis is approximate closure of a finite-dimensional Koopman operator, error bounds in the steepness parameter $$\mathrm{LGO}_{\text{hard}}(x;\alpha,\beta)=g(x;\alpha,\beta).$$9, and mesh resolution $\beta\to\infty$0 for approximating nonlinear vector fields (Johnson et al., 2017). Again, the common element is switch-like logistic structure, but the intended semantics differ: Koopman observables serve as a lifting basis, not as auditable threshold operators.

Taken together, these neighboring literatures suggest that logistic gating is a general mechanism for expressing regime dependence, while LGO in symbolic regression specializes that mechanism to typed threshold primitives, z-space learning, inversion to raw units, and anchor-based audit.