Behavior Learning (BL)¶
Conference: ICLR2026
OpenReview: https://openreview.net/forum?id=bbAN9PPcI1
Code: https://github.com/MoonYLiang/Behavior-Learning (pip install blnetwork)
Area: Interpretable Machine Learning
Keywords: Intrinsic Interpretability, Identifiability, Utility Maximization, Inverse Optimization, Energy-based Models
TL;DR¶
Inspired by behavioral science, this paper directly incorporates the assumption that "observations are solutions to an optimization problem" as a learnable module. Each module is a Utility Maximization Problem (UMP) expressible in symbolic form. These are hierarchically stacked into a composite utility function that induces a Gibbs distribution for prediction/generation, simultaneously achieving strong predictive power, intrinsic interpretability, and (in the IBL variant) parameter identifiability.
Background & Motivation¶
Background: Interpretable Machine Learning (ML) aims to fit complex phenomena while remaining transparent. Existing pathways to mitigate the "performance-interpretability tradeoff" generally fall into four categories: Additive Models (GAM/EBM/NAM), Concept Bottleneck Models, Rule/Scoring Systems, and Shape-Constrained Neural Networks.
Limitations of Prior Work: Most of these methods function as "interpretability add-ons" to existing ML methods, suffering from two fundamental issues. First is the mismatch with scientific theory—they do not originate from scientific modeling paradigms like optimization problems or differential equations, making it difficult to extract knowledge acceptable to the scientific community. Second is non-unique explanations—most models are non-identifiable, where the same prediction can correspond to multiple sets of parameters/explanations. This prevents reliable estimation of "true parameters" and reduces scientific credibility in the Popperian sense of falsifiability.
Key Challenge: High-performance models (Deep Networks) are opaque, while intrinsic interpretable models fail to capture complex non-linearities. Moreover, even when interpretable, the explanations may be non-unique and non-identifiable. To create "scientifically usable" interpretable ML, one must bind predictive power, intrinsic interpretability, and identifiability together.
Goal: Design a general framework that alleviates the performance-interpretability tradeoff, possesses scientific foundations (based on optimization), and ensures parameter identifiability.
Key Insight: The authors leverage the fundamental paradigm of behavioral science—Utility Maximization: the behavior of agents can be viewed as solving an optimization problem (UMP) to maximize subjective utility under constraints. Crucially, Theorem 2.2 states that any optimization problem with equality/inequality constraints can be equivalently rewritten as a UMP. Thus, a framework with UMPs as building blocks is inherently universal, applicable to scientific fields where "outcomes are optimization solutions" (e.g., macroeconomics, statistical physics, evolutionary biology), essentially performing data-driven inverse optimization.
Core Idea: Replace "black-box non-linear layers" with "learnable UMP modules." Each module can be written as a symbolic optimization problem. After hierarchical composition, a condition Gibbs (energy) distribution is used to model data, embedding interpretability into the structure rather than relying on post-hoc explanations.
Method¶
Overall Architecture¶
BL views the response \(y\) in a sample \((x,y)\) as the localized result of an agent solving interacting UMPs. The input consists of context features \(x\in\mathbb{R}^d\), and the response \(y\) can contain both discrete and continuous parts \((y_{disc}, y_{cont})\). The pipeline composes several learnable UMP modules \(B(x,y)\) into a composite utility function \(BL(x,y)\), which then parametrizes a conditional Gibbs distribution:
for prediction and generation. As the temperature \(\tau\to 0\), the distribution collapses to a Dirac measure at \(\arg\max_y BL(x,y)\), recovering the deterministic optimal response from the composite UMP. The entire network is trained end-to-end. By replacing penalty functions with smooth monotonic forms, the IBL variant ensures unique explanations and recovery of true parameters under mild conditions.
%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
A["Input (x, y)<br/>Context Features + Response"] --> B["UMP Module B(x,y)<br/>Utility-Inequality-Equality Penalty"]
B --> C["Hierarchical Composition<br/>Single / Shallow / Deep"]
C --> D["Conditional Gibbs Distribution<br/>BL(x,y) as Energy Function"]
D -->|"Cross-entropy for Discrete<br/>Score Matching for Continuous"| E["Prediction / Generation"]
C -.->|"Smooth monotonic penalties → IBL"| F["Identifiability<br/>Unique explanation + Parameter recovery"]
Key Designs¶
1. UMP Module: Making an optimization problem a learnable layer
This is the core mechanism to resolve the "mismatch with scientific theory." A standard UMP is \(\max_{y} U(x,y)\) s.t. \(C(x,y)\le 0,\ T(x,y)=0\), where \(U\) is subjective utility, \(C\) represents inequality constraints (resources), and \(T\) represents equality constraints (belief consistency or conservation laws). Using Theorem 2.1 (Local Exact Penalty Reconstruction), the constrained problem is rewritten as an unconstrained penalty form and parameterized as a learnable module:
where \(\phi\) is increasing, \(\rho(z)=\max\{z,0\}\) penalizes inequality violations, and \(\psi(z)=|z|\) penalizes equality deviations. The default instantiation is \(B=\lambda_0^\top\tanh(p_u)-\lambda_1^\top\mathrm{ReLU}(p_c)-\lambda_2^\top|p_t|\), where \(p_u,p_c,p_t\) are polynomial feature maps. The bounded \(\tanh\) corresponds to "diminishing marginal utility," while ReLU and \(|\cdot|\) provide soft penalties. Critically, each module can be mapped back to a symbolic UMP—the \(\tanh\) term is the objective, the ReLU term is the inequality constraint, and the absolute value term is the equality constraint. With polynomial bases, transparency is comparable to linear regression.
2. Hierarchical Composition: From single UMP to "Macro-Micro" optimization hierarchies
Since a single UMP has limited expressive power, \(B\) is used as a block for hierarchical composition. BL(Single) uses one \(B\) for maximum interpretability (direct symbolic UMP). BL(Shallow) stacks one or two layers, concatenating parallel \(B_{\ell,i}\) modules into a vector \(B_\ell(x,y)=[B_{\ell,1},\dots,B_{\ell,d_\ell}]^\top\) for the next layer. BL(Deep) extends beyond two layers:
Deep versions optionally use skip connections. This hierarchy represents "Coarse-graining/Renormalization" in science: bottom \(B\) blocks are micro-level primary preferences, aggregated layer-by-layer into macro-level tradeoffs and representative agents. Interpretability is thus bottom-up and traceable: Raw features → Micro-optimization blocks → Macro-aggregation/Coarse-grained constructs → Macro-optimization system.
3. Gibbs Distribution + Mixed Objective: Training composite utility as an energy function
Using \(BL(x,y)\) as an energy function, the conditional Gibbs distribution models the data, aligning "utility maximization" with "density maximization" as \(\tau\to 0\). Training targets are split: Cross-entropy for \(y_{disc}\), and Denoising Score Matching for \(y_{cont}\) to bypass the intractable partition function \(Z_\tau\). The final loss is a weighted sum:
The authors prove (Theorem 2.3) that BL (and IBL) possess the Universal Approximation property: with sufficient capacity, they can approximate any continuous conditional density in the KL sense.
4. IBL: Smooth monotonic penalties for Identifiability
To achieve "unique explanations," IBL imposes stricter constraints: \(\phi_{id},\rho_{id}\) are strictly increasing, and \(\psi_{id}\) is symmetric and strictly increasing about \(|\cdot|\), with all being \(C^1\). Instantiated as \(B_{id}=\lambda_0^\top\tanh(p_u)-\lambda_1^\top\mathrm{softplus}(p_c)-\lambda_2^\top(p_t)^{\odot2}\). Under Assumption 2.1 (injective parameter mapping + linear independence + canonical ordering), this yields: Identifiability (Theorem 2.4), Loss Identifiability (Theorem 2.5), Consistency (Theorem 2.6: \(\hat\theta_n\xrightarrow{p}\theta^\bullet\)), and Universal Consistency (Theorem 2.7: convergence to the true distribution in KL even under misspecification).
Key Experimental Results¶
Standard Prediction Tasks (10 datasets × 8 seeds)¶
Comparison across 10 baselines (NNs, Trees, Boosting, Bayesian, Linear Regression).
| Model | F1-Macro Avg Rank | Positioning |
|---|---|---|
| SOTA Black-box | Tier 1 | Performance Ceiling |
| BL(Shallow) | Tier 1/2 (No sig. diff from SOTA) | Best among Intrinsic Interp. Models |
| BL(Single) | Close Follower | Strongest Interpretability |
| MLP | Surpassed by BL(Shallow) | Black-box Control |
Key finding: BL reaches the first tier in AUC and F1-Macro, standing as the best-performing intrinsic interpretable model.
High-Dimensional Scalability (Image + Text, vs. E-MLP)¶
Depths \(d\in\{1,2,3\}\), aligned parameters, no skips.
| Dataset | Metric | E-MLP (d=3) | BL (d=3) |
|---|---|---|---|
| MNIST | OOD AUROC | 87.76 | 92.92 |
| Fashion-MNIST | OOD AUROC | 83.13 | 89.24 |
| MNIST | ID Acc | 98.14 | 97.93 |
| Fashion-MNIST | ID Acc | 89.33 | 88.79 |
BL shows superior OOD detection while maintaining comparable ID accuracy. BL also exhibits better calibration (ECE/NLL) on text datasets.
Key Findings¶
- BL provides a "downward shift" of the Pareto frontier: achieving performance comparable to black-box models while gaining transparency.
- Case Study (Boston Housing): A trained BL(Single) can be rewritten as a "representative buyer" UMP: \(p_u\approx(1-P)(1+P-RM)+\tilde R_u\). RM (rooms) dominates preferences, LSTAT (low income) enters budget constraints, and CRIM (crime) appears only in "belief" terms.
- BL(Deep) [5,3,1] recovers hierarchical patterns: 5 micro-preferences → 3 macro-tradeoffs → 1 representative buyer, aligning with economic literature and renormalization principles.
Highlights & Insights¶
- Inverse Optimization as Learnable Layers: Theorem 2.2 transforms any optimization into a UMP, making the network inherently scientific. This moves beyond post-hoc explanations to structural interpretability.
- Identifiability as a First-Class Citizen: Unlike most models that settle for "being interpretable," BL demands "unique explanations" and provides statistical guarantees through IBL.
- Unified Energy Perspective: The use of Gibbs distributions aligns utility maximization with probabilistic modeling, and denoising score matching makes it computationally feasible.
- Hierarchy as Coarse-graining: Interpreting deep architectures as micro-to-macro optimization tiers rather than just capacity stacks offers a powerful tool for scientific modeling.
Limitations & Future Work¶
- Strong Assumptions on Misspecification: Consistency in parameter recovery relies on the data being generated by some \(\theta^\star\), which is rarely true in practice.
- Approximate Symbolic Explanations: Converting polynomials to "readable UMPs" involves keeping only top coefficients; there is an inherent tradeoff between readability and fidelity.
- Interpretability Decay in Deep Models: BL(Deep) often uses affine maps for efficiency, reducing symbolic granularity to qualitative patterns.
- Small-to-Medium Experimental Scale: While robust on tabular data and basic images/text, validation on truly large-scale or multi-modal datasets is pending.
Related Work & Insights¶
- vs. GAM/CBM/Rule Models: These lacks scientific grounding and are often non-identifiable. BL builds on UMP optimization logic with identifiability guarantees.
- vs. Energy-based Models / E-MLP: While E-MLP is a black box, BL structures the energy function as a symbolic UMP composite, matching performance while adding transparency.
- vs. Neuro-symbolic / Symbolic Regression: BL constrains the symbolic structure to the "utility-constraint" paradigm and provides formal consistency theory.
Rating¶
- Novelty: ⭐⭐⭐⭐⭐ Transforming "Any Optimization = UMP" into layers with identifiability is highly novel.
- Experimental Thoroughness: ⭐⭐⭐⭐ Comprehensive across prediction and case studies, though lacking massive-scale validation.
- Writing Quality: ⭐⭐⭐⭐ Clear theory-method-experiment loop, though mathematically dense.
- Value: ⭐⭐⭐⭐⭐ Provides a theoretically grounded paradigm for "scientifically usable" interpretable ML.