[KongchangAI]
· 2 min read· 1,389 words

Farewell to Infeasible Decisions: Descriptive Learning Provides Feasibility Guarantees for Binary Decisions

Farewell to Infeasible Decisions: Descriptive Learning Provides Feasibility Guarantees for Binary Decisions

Descriptive learning directly optimizes feasible decision policies, guaranteeing feasibility and cutting costs under distribution shift.

The article addresses the shortcomings of predictive learning in binary decisions under uncertainty—lack of counterfactual knowledge and feasibility guarantees—and introduces a descriptive learning framework. It directly jointly optimizes a linear decision policy and feasible allocations, embeds feasibility into training, and extends to a Wasserstein ambiguity set-based distributionally robust chance-constrained form for a robustness–solution quality trade-off. In unit commitment and personnel assignment experiments, the robust model is more conservative when distributions match; under distribution shift, unmet demand drops sharply and total cost falls by up to 76.5% versus predictive learning. The article argues that in production environments with changing distributions, feasibility guarantees matter more than prediction accuracy, offering a new path for energy, transportation, and workforce fields.

In energy scheduling, traffic management, and workforce planning, binary decisions under uncertainty are ubiquitous. Traditional approaches rely on mixed-integer solvers, but the computational cost is high. In recent years, machine learning has been used to accelerate or even replace solvers, with the mainstream approach being predictive learning—training models to reproduce optimal decisions from labeled data. However, this paradigm has two fundamental flaws: the model does not understand counterfactual outcomes, and it cannot guarantee that the generated decisions are feasible. Under distribution shift, the problem is further amplified, and policies trained on historical data may systematically produce infeasible solutions.

To address this pain point, a new study proposes a descriptive learning framework that directly jointly optimizes over linear decision policies and feasible allocations, and extends it to a distributionally robust chance-constrained form, achieving a trade-off between robustness and solution quality while guaranteeing feasibility.

Descriptive learning framework diagram

Two Major Pain Points of Predictive Learning

Predictive learning transforms the decision problem into supervised learning: given scenario features, predict the optimal decision label. This approach performs well when data is abundant and the distribution is stable, but it is essentially imitating historical optimal solutions rather than understanding the feasibility conditions of decisions.

The first pain point is the lack of counterfactual knowledge. The model only sees "what was done," not "what would happen if a different decision were made." The second pain point is more critical: there is no feasibility guarantee. Predicted decisions may violate constraints and require post-hoc repair, and the repair process may destroy optimality. Under distribution shift, these two problems compound—the model may be feasible on the training distribution but may frequently violate constraints on the test distribution.

According to the paper, existing methods may produce systematically infeasible decisions under distribution shift, which is unacceptable in real-world systems.

Counterfactual knowledge refers to the model's understanding of "what the outcome would be if a different decision were taken." In predictive learning, the model only learns from historical data "what the optimal decision is in a given scenario," but never observes or evaluates the consequences of other feasible decisions. This leads to a causal confusion in which the model cannot distinguish between "the decision was adopted because it was optimal" and "the decision became the label because it was adopted." The lack of counterfactual knowledge makes it difficult for the model to adapt flexibly to new scenarios, and it tends to mechanically replicate historical patterns, making it especially vulnerable under distribution shift. Descriptive learning bypasses the reliance on counterfactual labels by directly optimizing the decision policy and its operational performance, thereby implicitly evaluating the consequences of different decisions during training.

Core Idea of Descriptive Learning

The framework abandons the route of "predicting optimal labels" and instead directly optimizes the decision policy. Specifically, it jointly optimizes a linear decision policy and feasible allocations across all training scenarios, so that the training process directly considers feasible alternatives and their operational performance.

The advantage of this approach is that the model no longer needs to learn a possibly nonexistent "optimal label," but instead learns a mapping from scenarios to decisions that is constrained to be feasible during training. Because the decision policy is linear, the optimization problem remains tractable, while feasibility is embedded in the learning objective.

Compared with predictive learning, descriptive learning turns feasibility from a "post-hoc check" into an "ex-ante guarantee," which is a fundamental paradigm shift.

A linear decision policy is a rule that maps scenario features (such as demand forecasts and price signals) to decision variables through a linear function, i.e., decision = linear coefficients × features + intercept. Compared with black-box models, linear policies have the advantages of strong interpretability and good optimization tractability: during training, policy parameters and feasible allocations can be jointly formulated as a linear or convex optimization problem, avoiding nonconvex training difficulties. At the same time, linear policies are naturally suited to being embedded in chance-constrained and robust optimization frameworks, because feasibility conditions can be transformed into constraints on the linear coefficients. The authors' choice of linear policies is a trade-off between expressiveness and computational feasibility, ensuring that the descriptive learning framework can guarantee feasibility during training without losing practical deployment value due to excessive model complexity.

Distributionally Robust Chance-Constrained Extension

To cope with distribution shift, the authors further extend the framework to a distributionally robust chance-constrained form using a Wasserstein ambiguity set. This form requires that, under the worst-case distribution within the ambiguity set, the decision satisfies feasibility with a predetermined probability.

This introduces a tunable trade-off: the stronger the robustness, the lower the solution quality may be, and vice versa. In practice, decision-makers can adjust the conservativeness of the constraints according to their risk preferences.

According to the paper's experiments, when the test distribution matches the training distribution, the robust model is more conservative; under distribution shift, the robust model significantly reduces unmet demand and lowers total cost by up to 76.5% relative to classical predictive learning.

A Wasserstein ambiguity set is a mathematical tool for characterizing distributional uncertainty. Centered on a reference distribution (usually the empirical distribution of the training data), it includes all distributions whose Wasserstein distance from the reference distribution does not exceed a given radius. The Wasserstein distance measures the minimum cost of "transporting" one distribution into another and can reflect the geometric differences between distributions. In distributionally robust optimization, the decision-maker requires the chance constraint to hold under all possible distributions within the ambiguity set, thereby obtaining immunity to distribution shift. The radius of the ambiguity set controls the degree of conservativeness: the larger the radius, the more distributions are considered, the stronger the robustness but the lower the solution quality may be; the smaller the radius, the closer it is to an ordinary chance constraint. This tool enables the descriptive learning framework to cope with the risk of mismatch between training and test distributions in a tunable manner.

Experimental Validation: Unit Commitment and Personnel Assignment

The paper validates the framework on two typical cases: the unit commitment problem and the personnel assignment problem. Both are typical binary decision scenarios involving discrete choices such as starting/shutting down generating units or assigning employees.

Out-of-sample experiments cover both cases where the test distribution matches the training distribution and where it does not. The results show that the robust model is more conservative when the distributions match, which is the expected behavior of robust optimization; under distribution shift, its advantages are obvious—unmet demand is greatly reduced and total cost is significantly lowered.

This result shows that the descriptive learning framework not only provides theoretical feasibility guarantees but also outperforms predictive learning on practical operational metrics, especially in environments with distributional uncertainty.

Implications for Practical Applications

For decision-makers in energy, transportation, workforce, and other fields, this study sends a clear signal: relying solely on predictive learning to replace solvers carries risks. In production environments where distributions may change, feasibility guarantees are more important than prediction accuracy.

The descriptive learning framework offers a new path: embedding feasibility constraints directly into the learning process and addressing distributional uncertainty through robust optimization. Although the computational complexity may be higher than that of simple predictive models, considering the operational losses caused by infeasible decisions, such investment is worthwhile.

In the future, the framework can be extended to more complex decision structures or combined with other robust optimization techniques to further enhance practicality.

One-line Summary

Descriptive learning directly optimizes feasible decision policies, guaranteeing feasibility and significantly reducing costs under distribution shift.

Paragraph Summary

The article addresses the shortcomings of predictive learning in binary decisions under uncertainty—namely, the lack of counterfactual knowledge and feasibility guarantees—and introduces a descriptive learning framework. The framework directly jointly optimizes a linear decision policy and feasible allocations, embeds feasibility into the training process, and extends to a distributionally robust chance-constrained form based on a Wasserstein ambiguity set, achieving a trade-off between robustness and solution quality. In unit commitment and personnel assignment experiments, the robust model is more conservative when distributions match; under distribution shift, unmet demand is greatly reduced, and total cost is lowered by up to 76.5% relative to predictive learning. The article points out that in production environments where distributions may change, feasibility guarantees are more important than prediction accuracy, and descriptive learning offers a new path for energy, transportation, workforce, and other fields.

#descriptive learning#predictive learning#distributionally robust optimization#chance constraints#Wasserstein ambiguity set#binary decisions#unit commitment#personnel assignment#feasibility guarantee#distribution shift
Share:

Related articles