Machine Learning–Guided Rule Selection

• Machine learning–guided rule selection is a data-driven framework that combines optimization techniques with expert input to derive interpretable decision rules.
• It employs methods such as submodular maximization, sequential covering, and reinforcement learning to balance accuracy with rule complexity.
• Applications span high-performance computing, medical decision support, and neuro-symbolic reasoning, enabling scalable and robust system designs.

Machine learning–guided rule selection refers to a suite of frameworks and algorithms in which machine learning techniques are leveraged to induce, select, or refine rules that drive decision-making in complex domains. Rules can be logical, probabilistic, or algorithmic; they may encode heuristics for algorithm selection, program optimization, data classification, or interpretable policy definition. Unlike fully manual or domain-expert–only approaches, these frameworks exploit data-driven optimization and statistical regularities, but may also incorporate expert input, interpretability, or domain constraints.

1. Conceptual Foundations and Problem Settings

The central purpose of machine learning–guided rule selection is to automate or assist in the derivation of actionable, high-impact, or interpretable rules. The core settings include, but are not limited to:

• Design-space exploration: Automatically discovering performant operation orderings, synchronizations, and resource bindings in parallel/distributed software (e.g., CUDA+MPI programs) (Pearson et al., 2022).
• Rule induction for prediction: Learning logical or statistical hypotheses (e.g., IF–THEN rules, DNF/CNF rule sets, decision lists) for classification, regression, regression, or survival analysis from labeled data (Sikora et al., 2018, Yang et al., 2022, Pellegrina et al., 2024).
• Meta-rule selection for logical induction: Learning to prune or prioritize higher-order meta-rule sets in the context of symbolic reasoning or visual abductive learning, to boost tractability and accuracy (Jin et al., 9 Mar 2025).
• Algorithm/policy selection: Mapping instance features to choices over a discrete set of execution rules, such as algorithmic pivot strategies, with the goal of optimizing empirical performance (Adham et al., 2021).
• Online correction or personalization of rules: Incrementally adapting heuristic system rules in streaming or drift-prone environments (Halford et al., 2020, Valente et al., 2021).
• Evolving interpretable classifiers in constrained domains: Using evolutionary and grammar-guided optimization to synthesize rule-based classifiers with high human interpretability (Torre-López et al., 28 Sep 2025).

These applications unify around the need to combine data-driven optimization, constraint handling, and interpretability or domain knowledge in the rule selection process.

2. Formalizations and Optimization Objectives

Machine-learning–guided rule selection typically formalizes rule learning and selection as a constrained optimization problem over a combinatorial hypothesis space. Common formal elements include:

• Rule set or policy representation: Rules are encoded as conjunctions, disjunctions, or ordered lists of literals (Boolean or numeric conditions), decision trees, or as symbolic logic programs. Rule sets may be unordered (DNF, as in (Yang et al., 2022)) or ordered (decision lists (Pellegrina et al., 2024)).
• Objective functions: Objectives balance predictive accuracy (often empirical risk on a validation set) with interpretability metrics (rule set size, literal count, overlap), and sometimes domain or sparsity penalties. For rule sets as DNF, a typical formal goal is

$$\min_{S \subseteq 2^{[d]}} \sum_{i=1}^n l(P_S(x_i), y_i) + \Omega(S)$$

where $l$ is a surrogate loss (see (Yang et al., 2022)), and $\Omega$ penalizes rule complexity.

• Approximation and generalization guarantees: Leveraging the VC-dimension of the rule class (e.g., for rule lists, $O(k z \log(d/z))$) enables uniform convergence results and sample-complexity bounds for provably near-optimal solutions via sub-sampling (Pellegrina et al., 2024).
• Optimization frameworks: Approaches include greedy (submodular maximization), coordinate descent (elastic net/logistic regression on rule binary indicators), evolutionary/grammar-based search, and reinforcement learning over meta-rule policies.

Design-specific cost objectives—such as execution time of traversals in CUDA+MPI DAGs (Pearson et al., 2022) or induction reward in logic metarule selection (Jin et al., 9 Mar 2025)—augment standard empirical risk minimization.

3. Representative Algorithmic Approaches

Submodular Optimization for Rule Sets:

Learning DNF rule sets is cast as a submodular maximization subject to cardinality and complexity constraints. By exploiting the submodularity of coverage functions and designing difference-of-submodular (DS) surrogate objectives, efficient greedy and local-search-based algorithms can construct compact high-accuracy rule sets with explicit approximation guarantees (Yang et al., 2022).

Sequential Covering and Guided Induction:

Sequential covering (separate-and-conquer) constructs rules one-by-one, greedily growing rules to maximize a quality metric (e.g., conditional entropy or log-rank, depending on the task), and pruning superfluous conditions. User preferences (required/forbidden attributes or conditions, initial rule seeds, maximum counts) are incorporated into the search, producing domain-aligned and more stable rule sets (Sikora et al., 2018). The pseudo-code in the cited work details the Grow, Prune, and GuidedGrow procedures.

Sampling-Based Approximation for Rule Lists:

Given the NP-hardness of globally optimal rule list learning, provably accurate approximations are constructed by sampling a subset of data points whose size is determined by the VC-dimension analysis. An exact solver (e.g., branch-and-bound) is applied to the sample, producing solutions with high-probability $(\epsilon, \theta)$-approximation guarantees for the risk on the full dataset (Pellegrina et al., 2024).

Grammar-Guided Evolutionary Rule Mining:

To enforce syntax validity and ensure interpretability, candidate rules are generated via context-free grammars that define the allowable structure (e.g., conjunctions of textual or bibliometric conditions). Genetic programming operates on derivation trees, evaluates fitness as a function of class-specific coverage/penalty, and maintains archives for candidate rules. The method efficiently produces small rule sets with transparent logical forms and competitive balanced-accuracy (Torre-López et al., 28 Sep 2025).

Meta-Rule Policy Optimization in Abductive Reasoning:

When meta-rule selection (searching a large set of logic templates) is computationally intractable, an attention-based selection policy is pre-trained on symbolic data. The policy maps symbolically-embedded cases and rules to a context-dependent mask over candidate meta-rules, drastically reducing the search space for logic abduction. Optimization is performed by policy-gradient RL (e.g., PPO) under a reward function tied to successful rule induction (Jin et al., 9 Mar 2025).

ML-Augmented Rule Selection for Discrete Algorithmic Choices:

Instance-based rule selection for algorithm portfolios is formulated as multi-class classification or multi-output regression, with features engineered for domain specificity (problem graph statistics, SVD decompositions, degree sequences). Decision forests or neural networks perform empirical cost or accuracy minimization for discrete choices (e.g., pivot rules, shortest-path algorithms) (Adham et al., 2021).

4. Integration with Other Learning and Inference Components

Rule selection is often embedded in wider neuro-symbolic or system pipelines, interacting with both statistical learners and symbolic components:

• Rule weighting and probabilistic reasoning: In hybrid systems like RLIE, LLM-generated rules are weighted by logistic regression (with elastic net regularization), and only the global combiner delivers robust inference; LLM-internal reasoning degrades with increased complexity or quantity of weights (Yang et al., 22 Oct 2025).
• Rule-based filtering for prediction: In interactive theorem proving, ILP-learned tactic applicability rules are used to filter or re-rank candidates from non-symbolic $k$-NN predictors, leading to significant gains in predictive accuracy and interpretability (Zhang et al., 2024).
• Personalized aggregation: Predicting, for each sample and rule, the probability that a rule will be correct enables patient-specific or instance-specific personalized aggregation, moving beyond unweighted voting in ensemble-induced rule sets (Valente et al., 2021).
• Online correction: Streaming or online adaptation of rule correction factors (e.g., selectivity in PostgreSQL) is achieved through stochastic, lightweight regression (linear, FM, shallow NN), handling concept or workload drift with a fixed learning rate or exponential forgetting (Halford et al., 2020).

5. Practical Applications and Case Studies

Machine-learning–guided rule selection demonstrably impacts diverse research and engineering domains:

• High-Performance Computing: Design rules for CUDA+MPI codes are automatically inferred through guided search and decision-tree analysis of sequence features, directly informing code revisions for performance gains (Pearson et al., 2022).
• Systematic Literature Reviews: Evolutionary grammar-based rule induction provides interpretable, high-recall screening classifiers for primary study selection, outperforming SVM-based baselines in both accuracy and interpretation (Torre-López et al., 28 Sep 2025).
• Medical Decision Support: Sparse, personalized rule sets constructed via ensemble extraction, LASSO selection, and correctness-modeling approach the accuracy of unconstrained forests while being an order of magnitude smaller. Weighted aggregation consistently raises AUC over baseline voting (Valente et al., 2021).
• Discrete Algorithm Portfolios: ML-driven rule selectors outperform fixed expert policies in choosing linear programming pivot strategies and all-pairs-shortest-paths solvers, with empirical gains on both instance- and distribution-level metrics (Adham et al., 2021).
• Neuro-symbolic Reasoning: RLIE demonstrates that LLM-driven rule hypothesis generation, when filtered and weighted by sparse logistic regression, yields interpretable, accurate, and calibration-preserving final models, with iterative improvement via feedback-driven refinement (Yang et al., 22 Oct 2025).
• Visual Generative Abductive Reasoning: Meta-rule policy learning achieves near-expert selection of logic templates in visual symbolic tasks, drastically improving the tractability and quality of the abduction process (Jin et al., 9 Mar 2025).

6. Interpretability, Constraints, and Limitations

While machine-learning–guided rule selection provides mechanisms to balance accuracy, interpretability, and domain constraints, these trade-offs remain central:

• Rule set size and complexity: Penalization schemes (e.g., $\ell_1$-regularization, modular penalties in submodular frameworks, grammar depth constraints) are critical for producing small, human-understandable models (Yang et al., 2022, Pellegrina et al., 2024, Torre-López et al., 28 Sep 2025).
• Feature and condition preferences: User- and domain-guided frameworks allow specification of included/excluded features and conditions, which can increase stability and domain alignment, often improving support and precision (as with GuideR (Sikora et al., 2018)).
• Transferability and robustness: Many rule sets, especially for program optimization or symbolic abduction, may be platform- or dataset-specific and require re-derivation when confronted with new architectures or semantically distant inputs (Pearson et al., 2022, Jin et al., 9 Mar 2025).
• Scalability: Sampling-based and evolutionary methods address the combinatorial explosion in large-scale applications, often combining theoretical sample- or complexity-bounds with practical pruning or fast local search (Pellegrina et al., 2024, Torre-López et al., 28 Sep 2025).
• Limitations on semantic reasoning: While LLMs can propose or refine rules, probabilistic score integration and calibration are most reliably delegated to linear or classical statistical models (Yang et al., 22 Oct 2025).

7. Comparative Summary Table

Framework	Rule Representation	Learning Principle	Key Domain
Submodular Maximization	DNF (OR-of-ANDs)	Greedy/DS decomposition	Classification
GuideR	IF–THEN, user-guided	Sequential covering+guidance	General
Sampling + Exact Solver	Ordered rule lists	VC-dimension-based sampling	Rule lists
Grammar-guided GP	Grammar-constrained CARs	Evolutionary (G3P)	Literature SLR
RLIE	Natural-language, weighted	LLM + logistic regression	Neuro-symbolic
Meta-rule Selection Policy	Meta-rules (attention mask)	Attention policy + RL (PPO)	Abductive logic
Tactic Prediction ILP	Horn-clauses (ILP)	ILP from enriched AST featur.	Theorem proving

This table summarizes the core frameworks and their distinguishing algorithmic and representational characteristics, highlighting the diversity of both domains and technical approaches across the field.

---

Machine learning–guided rule selection synthesizes combinatorial optimization, statistical learning, expert knowledge integration, and computational efficiency to produce rules that are not only accurate and robust but also interpretable and actionable in complex real-world settings. The continued development of theoretically-founded, scalable, and domain-aligned frameworks remains an active research frontier.