Affine Decision Rules in Stochastic Optimization

• Affine Decision Rules (ADR) are tractable policy parameterizations that express multistage decisions as affine functions of past uncertainties, ensuring non-anticipativity.
• ADRs transform infinite-dimensional robust and stochastic optimization problems into finite-dimensional linear or conic programs, improving computational feasibility and structural simplicity.
• Extensions like piecewise-affine and quadratic decision rules enhance approximation accuracy at higher computational cost, as shown in applications like scheduling and inventory control.

An affine decision rule (ADR) is a tractable policy parameterization for multistage stochastic and robust optimization, where decisions are restricted to be affine functions of the observed uncertain parameters. This approach enables the practical solution of otherwise intractable adjustable robust and stochastic programming problems by projecting the infinite-dimensional policy search onto a finite-dimensional space of affine mappings. ADRs balance computational tractability, theoretical interpretability, and structural simplicity, and serve as the canonical baseline for more expressive classes such as piecewise-affine and quadratic decision rules.

1. Mathematical Formulation and Scope

Given a $T$-stage optimization problem under uncertainty, an affine decision rule takes the form

$x_t(\xi^{\,t-1}) = A_t\,\xi^{\,t-1} + b_t$

where $x_t$ is the stage-$t$ decision, $\xi^{\,t-1} = (\xi_1,\ldots,\xi_{t-1})$ are the revealed uncertain parameters up to stage $t{-}1$, $A_t$ and $b_t$ are the rule coefficients, and the affine dependence is with respect to realized history. This structure adheres strictly to the non-anticipativity principle: the mapping at time $t$ depends only on the information $\mathcal F_{t-1}$ available up to $x_t(\xi^{\,t-1}) = A_t\,\xi^{\,t-1} + b_t$0 (Thomä et al., 2024, Bodur et al., 2017).

In robust, stochastic, and distributionally robust settings, ADR-policies replace the infinite-dimensional mapping $x_t(\xi^{\,t-1}) = A_t\,\xi^{\,t-1} + b_t$1 with an explicit parametric form. This reduction enables the recasting of the original problem—often infinite-dimensional and nonconvex—into a tractable linear or conic program with respect to the ADR parameters (Woolnough et al., 2020, Bodur et al., 2017).

2. ADRs in Adjustable Robust and Stochastic Optimization

ADRs emerged as a foundational approximation in Adjustable Robust Optimization (ARO) with convex uncertainty sets. For a two-stage robust linear program with ellipsoidal uncertainty, the affine response is

$x_t(\xi^{\,t-1}) = A_t\,\xi^{\,t-1} + b_t$2

where $x_t(\xi^{\,t-1}) = A_t\,\xi^{\,t-1} + b_t$3 is the realized disturbance (subject to, e.g., $x_t(\xi^{\,t-1}) = A_t\,\xi^{\,t-1} + b_t$4), and $x_t(\xi^{\,t-1}) = A_t\,\xi^{\,t-1} + b_t$5, $x_t(\xi^{\,t-1}) = A_t\,\xi^{\,t-1} + b_t$6 are the optimization variables. The uncertain constraint

$x_t(\xi^{\,t-1}) = A_t\,\xi^{\,t-1} + b_t$7

is rendered explicit by the affine parametrization and may be equivalently reformulated via robust optimization theory (e.g., S-lemma, strong duality) as a semidefinite or second-order cone program (SDP/SOCP) (Woolnough et al., 2020).

For multistage stochastic linear programs, LDRs (linear decision rules—an equivalent term) yield an explicit static LP if data are deterministic and the support $x_t(\xi^{\,t-1}) = A_t\,\xi^{\,t-1} + b_t$8 is a polyhedron. All policy variables become affine in stagewise uncertainty, and the expected cost reduces to a function of the ADR parameters and corresponding robust or stochastic moment matrices (Bodur et al., 2017).

3. Extensions: Piecewise-Affine, Quadratic, and Hybrid Rules

The limited flexibility of ADRs in capturing strongly nonlinear or discontinuous policy structures motivates several extensions:

• Piecewise-Affine Decision Rules (PADRs): The policy is constructed by partitioning the uncertainty domain into polyhedral cells and assigning an affine rule per cell; this can be achieved via systematic lifting, folding, and convexification procedures (Thomä et al., 2024, Zhang et al., 2023). Enhanced “lift & tighten” algorithms further improve stochastic and distributionally robust performance by tighter outer approximations in the lifted space.
• Quadratic Decision Rules (QDRs): Policies of the form $x_t(\xi^{\,t-1}) = A_t\,\xi^{\,t-1} + b_t$9 capture quadratic nonlinearity. The same mathematical program as for ADRs admits exact SOCP or SDP representations when restricted to QDRs—setting $x_t$0 yields the ADR as a special case. QDRs generally dominate ADRs in policy quality at increased computational cost (Woolnough et al., 2020).
• Hybrid and Two-Stage LDRs: Imposing ADRs on select variables/stages—e.g., only on state variables—and leaving others flexible reduces conservatism. Two-stage LDRs typically yield dramatically improved upper-bound policies in MSLPs (Bodur et al., 2017).

4. Algorithmic and Computational Properties

ADRs enable explicit transformations of robust or multistage programs to convex programs:

• SDP/SOCP Reformulations: For ellipsoidal and box-uncertainty, ADR-formulated two-stage robust programs can be reduced to SDPs or SOCPs, retaining exactness with polynomial-time solvability. There is no approximation gap between the reformulated ADR-SDP/SOCP and the original robust counterpart (Woolnough et al., 2020).
• LP Approximations: In continuous-time robust scheduling, affine-in-current-demand rules reduce infinite-dimensional optimization over function spaces to a finite LP in the rule coefficients. Cutting-plane methods efficiently handle the robust constraint sets exploiting their polyhedral structure (Cho et al., 1 Apr 2025).
• Sample Average Approximation (SAA): For stochastic or data-driven programs, SAA reduces ADR-approximated problems to tractable LPs. Monte Carlo simulation enables statistical upper and lower bounds on the true value (Bodur et al., 2017).

A sample computational benchmark is summarized below.

Model	N (problem size)	ADR solve time (s)	QDR solve time (s)	ADR realized gap (%)	QDR realized gap (%)
Lot-sizing	2	0.02	0.09	67.7	64.7
Lot-sizing	8	0.18	46.5	71.9	64.2

QDRs achieve consistent improvements at increased computation (Woolnough et al., 2020).

5. Theoretical Properties and Approximation Quality

ADR-based policies yield provable upper (primal) or lower (dual) bounds on optimal value functions of multistage programs. In static LDRs (affine rules applied globally), the bound is generally loose; two-stage or hybrid LDRs reduce the gap substantially (Bodur et al., 2017).

Expressiveness of ADRs is fundamentally limited: any deterministic $x_t$1-Lipschitz function can be approximated to error $x_t$2 by a $x_t$3-piece PADR, while the ADR (i.e., $x_t$4) may suffer large approximation error in strongly nonlinear settings (Zhang et al., 2023).

Empirically, in stochastic inventory or newsvendor problems, PADRs dramatically outperform ADRs both in out-of-sample cost and statistical efficiency, especially under strong nonlinearity and higher-dimensional covariates (Zhang et al., 2023, Thomä et al., 2024).

6. Applications and Empirical Evidence

ADRs are deployed in continuous-time robust generation scheduling under demand uncertainty: the affine-in-current-demand rule ensures non-anticipativity, feasibility, and reduces to an LP solved in a handful of robust cuts, with all tested random demand trajectories respecting system constraints (Cho et al., 1 Apr 2025). In multistore lot-sizing, ADR SOCPs solve in under a second for real-world instance sizes. For multi-period inventory control, static LDRs yield $x_t$5–$x_t$6 optimality gaps, which are reduced to below $x_t$7 by two-stage LDRs (Bodur et al., 2017).

Scenario-based nonparametric rules, in contrast, can catastrophically violate feasibility constraints on fresh trajectories, while ADRs, by construction, maintain robust feasibility (Cho et al., 1 Apr 2025).

7. Limitations, Alternatives, and Research Directions

ADRs, while computationally attractive and providing explicit guarantees, are inherently limited in approximation power due to their global linearity. In robust programs, extending ADRs to classical piecewise-affine forms offers no gain unless the lifting-and-tightening approach is used (Thomä et al., 2024). QDRs and hybrid multi-stage ADRs consistently improve solution quality but at the cost of more complex conic formulations and longer solve times (Woolnough et al., 2020). In high-dimensional stochastic settings, PADR-based empirical risk minimization dominates both ADR and kernel-based methods in sample complexity and robustness (Zhang et al., 2023).

Current research focuses on tractable outer approximation schemes, improved cut-generation algorithms, and integrating data-driven constraints (via Wasserstein ambiguity sets) within the ADR/PADR paradigm (Thomä et al., 2024). The NP-hardness of grid-cut separation in PADRs motivates further study of efficient algorithms for high-dimensional policy design.

---

References:

• (Cho et al., 1 Apr 2025) Robust Continuous-Time Generation Scheduling under Power Demand Uncertainty: An Affine Decision Rule Approach
• (Thomä et al., 2024) A Note on Piecewise Affine Decision Rules for Robust, Stochastic, and Data-Driven Optimization
• (Zhang et al., 2023) Data-driven Piecewise Affine Decision Rules for Stochastic Programming with Covariate Information
• (Woolnough et al., 2020) Exact Conic Programming Reformulations of Two-Stage Adjustable Robust Linear Programs with New Quadratic Decision Rules
• (Bodur et al., 2017) Two-stage Linear Decision Rules for Multi-stage Stochastic Programming