Decision and Optimization Mapping

• Decision and Optimization Mapping is a framework that translates decision predictions into structured, mathematical optimization models using diverse paradigms.
• It integrates classical multi-stage stochastic models, decision-dependent robust methods, and learning-based approaches to optimize performance under uncertainty.
• The methodology leverages end-to-end machine learning, influence diagrams, and dynamic programming to achieve efficient, scalable, and interpretable decision systems.

Decision and Optimization Mapping refers to the systematic translation of decision-related information or predictions into the structure and solution of mathematical optimization models. The mapping encompasses a range of paradigms, from classical multi-stage stochastic optimization and combinatorial approaches to modern, machine learning–integrated frameworks that optimize the final decision quality directly. These methodologies now address not only static or parameterized problems, but also extend to scenarios where the optimization itself influences the environment or downstream data. Approaches to decision and optimization mapping vary by problem structure (linear, discrete/combinatorial, nonlinear), information acquisition (static, sequential, decision-dependent), degree of learning/automation, and treatment of uncertainty and feedback.

1. Classical and Sequential Decision-Optimization Paradigms

Traditional decision and optimization mapping centers on constructing explicit models that relate decisions to outcomes through maximization or minimization of an objective over a feasible set, potentially under uncertainty or multi-stage information revelation.

Multi-Stage Influence-Diagram MILPs: Multi-stage stochastic or robust optimization with complex dependencies between decisions and uncertain parameters can be encoded as influence diagrams, which are then automatically compiled into mixed-integer linear programs (MILPs). Each node in the diagram models a decision or chance event; edges represent informational or causal structure. Variables encode local policies, path probabilities, scenario reward, and logical or risk constraints. The MILP formulation guarantees global optimality over all feasible strategies, supports Pareto frontier enumeration in multi-objective cases, and accommodates non-recall, resource, and probabilistic constraints (Salo et al., 2019).

Decision-Dependent Robust and Distributionally Robust Optimization: When the timing or scope of information acquisition is determined by the decision-maker (e.g., through costly probes or staged investments), the information structure must be dynamically mapped to constraints and uncertainty sets. Decision-dependent robust frameworks use explicit mapping from probing decisions to information sets. This mapping enters the optimality equations as modified feasible sets dependent on prior information decisions. K-adaptability and its extensions approximate dynamic-contingency recourse by a tractable enumeration of policy candidates, yielding scalable robust MIPs for problems such as preference elicitation, portfolio design, and resource allocation (Vayanos et al., 2020). Decision-dependent distributionally robust optimization (DRO) maps the decision vector into the ambiguity set’s center and radius, yielding convex reformulations that incorporate decision sensitivity in the risk set geometry (Fonseca et al., 2023).

Dynamic Distribution and Feedback Effects: When decisions influence not only the immediate reward but also the environment or future data, the optimization must anticipate endogenous distribution shifts. The mapping is encoded via contraction dynamics mapping decisions to the invariant distribution, and the optimization targets steady-state expectations under this coupled measure. Online algorithms for such problems must align decision updates with gradient information that propagates through the distributional feedback, leading to stationarity and generalization rates that depend explicitly on contraction and Lipschitz properties of the feedback mapping (He et al., 10 Mar 2025).

2. Learning-Based Predict-Then-Optimize and Decision-Focused Learning

A prominent class of recent work addresses the setting in which certain problem parameters are not directly observable and must be predicted from contextual features prior to optimization.

Predict-Then-Optimize (PtO): The standard pipeline first learns a mapping (e.g., neural net, tree) from features to unknown problem parameters (e.g., cost coefficients), then solves the resulting deterministic optimization. However, optimizing for prediction accuracy alone (mean squared error in parameters) often yields high decision regret—the difference in true objective value between the predicted and optimal parameterization (Kotary et al., 2023).

Decision-Focused Learning (DFL): DFL aligns the predictive loss with the downstream decision quality, typically by minimizing decision regret rather than parameter error. This introduces a computational challenge: the loss function includes an argmin over the optimization problem, necessitating solver calls within each training iteration. For linear programs, specialized solvers can permit differentiation via the KKT system or similar sensitivity tools. In combinatorial settings, non-differentiability is addressed by learning-to-rank surrogates or contrastive losses over sampled solution subsets (Mandi et al., 2021).

Solver-Free Surrogates for LPs: For linear optimization, geometric properties of polytopes are leveraged to construct surrogate losses. The optimality condition that the objective at an optimal extreme point must not be exceeded by that at any neighboring vertex supports the construction of a piecewise-linear hinge loss over adjacent vertices. Training proceeds by precomputing the adjacency set and replacing all solver calls in the loop with fast vector operations. This produces convex, efficiently computable losses whose minimization drives decision regret to zero when satisfaction of all adjacent inequalities is achieved (Berden et al., 28 May 2025).

Framework	Structure	Mapping Mechanism	Optimization Phase
Classical influence diagrams	Discrete multi-stage	Decision policy → scenario paths	MILP for all feasible policies
Decision-dependent DRO/RO	Distributional/robust	Decision x → ambiguity/information sets	Convex reformulation/MILP
Data-driven DRO (Pareto principle)	Data-to-decision	Statistic T(D) → ambiguity set (LDP-based)	Single robust program
DFL (general)	Parametric/comb.	Features x → parameter → decision	Gradient updates through structure
Solver-free DFL for LPs	LPs	Feature mapping → cost vector → adjacents	Hinge loss over adjacent vertices

3. Integration with Machine Learning: End-to-End and Proxy Models

Recent developments integrate optimization directly into the ML pipeline, or bypass parametric prediction entirely in favor of feature-to-decision surrogates.

Joint Learning to Optimize from Features (LtOF): Instead of predicting parameters θ and solving for y*, one directly learns a mapping from inputs x to the (nearly) optimal decision y*(θ(x)), using a regret or feasibility-penalized loss and architectures such as neural networks, with constraint satisfaction enforced by projected gradient or primal-dual training. This paradigm eliminates the need for solver differentiation, reduces inference latency, and provides competitive or superior performance to both two-stage and end-to-end differentiable PtO pipelines, particularly in large-scale or nonconvex/quadratic settings (Kotary et al., 2023).

Proxy Submodular or Combinatorial Decision Layers: In problems where the decision is the solution to a submodular or combinatorial maximization (e.g., vehicle routing, coverage), submodular maximization is embedded as a differentiable layer using stochastic smoothing or entropy-regularized greedy methods. The ML model maps observed context to function parameters or decisions, and the soft, differentiable optimization layer ensures that gradients propagate through decision quality rather than predicted values, directly targeting final performance (Shi et al., 2023).

Interpretability via Structured Models and Decision Trees: Where interpretability or model parsimony is desired, feature-to-decision mappings are constructed as decision trees, with leaves split according to decision-quality-aware criteria such as the "SPO loss" (suboptimality in true objective). These trees provide translatable rules from observed features to prescribed actions, offering clear, transparent mappings from context space into the optimization action space (Elmachtoub et al., 2020).

4. Decision Diagram and Dynamic Programming–Based Mappings

In high-combinatorial or discrete spaces, decision diagrams (DDs)—layered, acyclic graphs that represent all feasible solutions—enable tractable mapping between decision sequences and their objective images.

Exact and Approximate Decision Diagrams: The solution space of a discrete optimization problem can be encoded as an exact DD, constructed by dynamic programming recursion, where paths from the root to the terminal node correspond one-to-one to feasible solutions. Intractability for large n is addressed by width-limited “relaxed” or “restricted” DDs that respectively over- or under-approximate the feasible set. The mapping from root-to-terminal path to decision vector defines a constructive, enumerative mechanism for mapping decisions into solution quality space, and the DD structure supports the extraction of upper and lower bounds, embedding in hybrid integer programming or constraint programming models, and the use of network-flow algorithms for primal-bound recovery (Castro et al., 2022).

Hybrid Integration into IP and CP: DDs can be embedded as flow-based constraints into mixed-integer or constraint programming models, providing either relaxations or global constraints that facilitate strong branching, faster convergence, or reduction in node counts for complex routing, scheduling, or network design applications.

5. Real-Time, Contextual, and Engineering Application Mapping

Certain applied optimization domains require context-to-decision mapping with strong real-time or domain-specific constraints.

Real-Time Decision Generators via Adversarial Learning: For contextually constrained continuous optimization where feasible sets and objectives change with features (e.g., personalized medical treatments, context-driven portfolio allocation), generative models are trained to map context to feasible, near-optimal decisions using adversarial barrier or interior-point inspired methods. Trained networks yield ms-scale inference latency and feasibility/optimality certificates are derived from probabilistic and empirical gap analyses (Babier et al., 2018).

Engineering-Informed Surrogate-Driven Optimization: In mechanical or physical systems (e.g., TBM operation), mappings are constructed by integrating physical models and data-driven estimation (dual-driven models), using constraints derived from first-principles or empirical rules and penalized at the surrogate loss level. The optimization phase relies on fast global enumeration over the compact decision surfaces, enabling principled, implementable mappings from machine or environmental measurements to explicit control decisions (Liu et al., 2022).

Sequential ML–Optimization Integration: For sequential combinatorial problems (multi-item lot-sizing, knapsack, facility location), encoder–decoder models with recurrent or attention-based structures produce predictions for period-by-period binary decisions. Infeasibility elimination is handled by tightness-prediction and adaptive variable-fixing, connecting ML predictions directly to a reduced-subproblem MIP, realizing several orders of magnitude speed-up with negligible suboptimality on large-scale, rolling-horizon problems (Yilmaz et al., 2023).

6. Challenges and Future Directions

Critical gaps and open questions persist throughout the decision and optimization mapping landscape:

• Model Alignment: Alignment of surrogate loss functions and the final operational objective is nontrivial. Many pipelines suffer when prediction loss is uncorrelated with post-optimization regret (Mandi et al., 2021).
• Scalability and Degeneracy: Strategies relying on local enumeration (e.g., adjacent vertex search for LPs, exact decision diagrams) can be challenged by exponential worst-case scaling in highly degenerate or large problems (Berden et al., 28 May 2025, Castro et al., 2022).
• Informativeness and Feedback: Correctly incorporating the structure of decision-dependent feedback—e.g., which pieces of information can be acquired, and when—is essential for robust dynamic programs, scenario-based optimization, and data-driven DROs (Vayanos et al., 2020, Sutter et al., 2020, He et al., 10 Mar 2025).
• Interpretability and Transparency: There is an increasing demand for mappings that are not only optimal but also explainable (decision trees, interpretable DDs), especially in regulated or stakeholder-facing domains.
• Dataset Construction and Benchmarking: The absence of large, validated corpora of narrative-to-optimization mappings, especially for end-to-end natural language interfaces or multi-modal problems, remains a bottleneck (Wasserkrug et al., 2024).
• Hybrid and Automated Reformulation: Embedding domain knowledge, solver selection, and dynamic optimization model improvement within ML-optimization pipelines remains a central technical challenge (Wasserkrug et al., 2024).

7. Summary Table: Representative Techniques

Mapping Paradigm	Decision --> Optimization Mechanism	Key Advantages/Limitations
Influence Diagrams → MILP (Salo et al., 2019)	Decision sequence encoding in acyclic graph, MILP	Full expressiveness, global opt
Decision-dependent ROB/DRO (Fonseca et al., 2023)	Decision parameterizes info set/ambiguity radius	Captures feedback, may be convex
PtO and DFL (LP/comb) (Berden et al., 28 May 2025, Mandi et al., 2021)	Parameter predictor or ranking loss, optimize for regret	End-to-end quality, heavy solve
Surrogate/LtOF (Kotary et al., 2023)	Direct x→decision mapping with constraint proxy loss	No solver calls, tunable gap
Decision Diagrams (Castro et al., 2022)	Exact/approximate enumeration via state-arc graphs	Compactness vs. bound strength
Engineering Surrogates (Liu et al., 2022)	Physics-informed DNNs as objective/constraint map	Real-system fidelity, case-tuned

Decision and optimization mapping thus encompasses a highly diverse spectrum of methods, each focused on structuring and exploiting the interplay between information, prediction, optimization, and decision-making under varying forms of uncertainty, scale, dynamics, and human interaction. These methodologies are foundational in enabling efficient, robust, and interpretable decision support in modern operations, engineering, and AI-driven application domains.