Two-Stage Co-Optimization (TSCO)

• Two-Stage Co-Optimization (TSCO) is a framework that splits complex decision problems into two interdependent stages, ensuring coordinated planning and adaptive recourse.
• The first stage typically sets initial decisions or global parameters, while the second stage refines solutions based on uncertainty, local feedback, or operational adjustments.
• Empirical studies show that TSCO enhances computational efficiency and solution performance across domains such as stochastic programming, energy systems, and autonomous planning.

Searching arXiv for the provided TSCO-related papers to ground the article in current records. arxiv_search_query: (Tan et al., 2021) TSCO two-stage optimization stochastic programming co-optimization arxiv_search_query: (Lou et al., 20 Jan 2025) nonlinear nonconvex two-stage optimization decomposition framework Two-Stage Co-Optimization (TSCO) denotes a family of optimization formulations and algorithms in which a coupled decision problem is organized into two coordinated stages rather than solved as a purely sequential pipeline or as an undifferentiated monolith. Across the literature, the two stages may correspond to first-stage planning and second-stage recourse under uncertainty, a cheap global stage followed by an accurate local refinement stage, or a day-ahead coordination stage followed by real-time operational adjustment. In all cases, the defining feature is not merely temporal separation, but explicit coupling: the first stage shapes, initializes, constrains, or prices the second stage, while the second stage supplies value information, feasibility information, or refinement feedback that determines the quality of the overall solution (Zhang et al., 26 Mar 2025, Tan et al., 2021, Cheng et al., 2019, Lou et al., 20 Jan 2025).

1. Conceptual scope and defining characteristics

TSCO is best understood as a structural pattern rather than a single canonical algorithm. In one major usage, it refers to two-stage stochastic optimization, where a first-stage decision is chosen before uncertainty is revealed and a second-stage recourse decision is chosen after realization of a random scenario. This pattern is written in the stochastic-programming literature as

$$\min_{x \in X \subseteq \mathbb{R}^{n_1}} f(x)=c(x)+E_{\omega}[h(x,\omega)], \qquad X=\{x: Ax=b,\ x\ge 0\},$$

with the recourse value obtained from a subordinate optimization problem in $y$ (Zhang et al., 26 Mar 2025). In another usage, TSCO describes algorithmic coarse-to-fine optimization: CobBO, for example, splits each round into a cheap global stage over the full $D$-dimensional space and a more accurate local stage over a selected low-dimensional coordinate subspace, with explicit transfer of means and variances from the first stage into the second (Tan et al., 2021). In energy systems, a two-stage transactive control framework separates day-ahead clearing-price forecasting from hourly rolling dispatch, again with the first stage providing the coordination signal used in the second (Cheng et al., 2019).

A common misconception is that any sequential two-step workflow qualifies as TSCO. The surveyed papers instead emphasize coordination, coupling, and information reuse. Conventional DNN/accelerator design, for example, is described as a “typical flow of the two-stage method” in which one first designs an application-specific DNN model with the highest accuracy and then builds an accelerator around that fixed DNN; YOSO criticizes this flow precisely because it is unidirectional and lacks hardware-to-network feedback (Chen et al., 2020). This suggests that TSCO, in the stronger sense used by recent work, requires the two stages to be jointly relevant to the same system-level objective rather than merely adjacent in a pipeline.

2. Mathematical archetypes

The most classical TSCO formalism is two-stage stochastic programming. In its generic form, the first-stage variable $x$ minimizes a deterministic cost plus an expected second-stage recourse term, while the recourse function itself is an optimization problem parameterized by $x$ and $\omega$: $$h(x,\omega)\triangleq \min_{y}\ g(y) \quad \text{s.t.}\quad Dy=\xi(\omega)-C(\omega)x,\quad y\ge 0.$$ The literature summarized here includes two-stage SQLP, with a quadratic first stage and linear second stage, and two-stage SQQP, with quadratic first and second stages (Zhang et al., 26 Mar 2025). A closely related decomposition form for nonlinear nonconvex TSCO writes

$$\min_{x\in\mathbb{R}^{n_0}} f_0(x)+\sum_{i=1}^N \hat f_i(x) \quad \text{s.t.}\quad c_0(x)\le 0,$$

with each second-stage value function defined by

$$\hat f_i(x)=\min_{y_i\in\mathbb{R}^{n_i}} f_i(y_i;x) \quad \text{s.t.}\quad c_i(y_i;x)\le 0.$$

Here the second-stage problems are scenario- or partition-specific subproblems evaluated independently for a given $x$ (Lou et al., 20 Jan 2025).

Another mathematical extension is two-stage optimization with recourse and revocation. In this model, first-stage commitments can later be partially revoked or, equivalently, reserved in stage 1 and activated in stage 2. The reservation-form objective for rooted Steiner tree is

$y$0

where $y$1 is the reserved first-stage set, $y$2 is the part actually used after the scenario is realized, and $y$3 is newly purchased recourse (Jiang, 2016). The key modeling novelty is the coupling constraint that the second-stage use of reserved objects must be a subset of what was reserved earlier.

A different but related archetype appears when TSCO is used to couple subsystems that were historically optimized sequentially. In launch-vehicle design, staging and trajectory are co-optimized in a single nonlinear program with decision vector

$y$4

where stage structural masses and phase-end pitch/yaw angles are optimized simultaneously under orbital-insertion and launch-safety constraints, including instantaneous impact point bounds (Ko et al., 2023). In autonomous driving, the integrated local-planning problem combines discrete lane decisions $y$5 and continuous trajectory variables $y$6 in one objective

$y$7

but is solved through a two-stage decomposition because the original problem is a mixed-integer nonlinear program with nonlinearity and nonconvexity (Liu et al., 2024).

3. Stage coupling mechanisms

The central technical question in TSCO is how information moves between stages. CobBO provides a particularly explicit answer. Stage 1 fits a computation-efficient kernel $y$8 on all observed full-space data to obtain a coarse posterior that “purposely smooth[s] away local fluctuations.” A coordinate block $y$9 is then selected, a subspace $D$0 is formed around a pivot point $D$1, and all past full-dimensional points are projected into that subspace through

$D$2

These projected locations become virtual points whose values and variances are estimated by Stage 1 and then supplied to Stage 2, where a richer kernel $D$3, typically ARD Matérn $D$4, is used for local Bayesian optimization (Tan et al., 2021). The resulting coupling is neither loose nor post hoc: the first-stage surrogate literally seeds the second-stage surrogate.

In nonlinear nonconvex TSCO, coupling often occurs through sensitivity information. The decomposition framework of (Lou et al., 20 Jan 2025) introduces an interior-point barrier smoothing of the second-stage problem so that the local second-stage solution behaves like a differentiable function of the first-stage parameters. If the barrier KKT Jacobian is nonsingular at a KKT point, the implicit function theorem yields a locally $D$5 solution map and a differentiable smoothed value function satisfying

$D$6

This allows the master problem to use gradients and Hessians derived from second-stage dual variables rather than from a nondifferentiable value function.

Surrogate-based TSCO replaces exact recourse evaluation by learned or metamodel approximations. One line of work alternates mixed-integer first-stage optimization with neural-network retraining so that the expected second-stage value function is approximated by a ReLU network and embedded into the first-stage MILP (Kronqvist et al., 2023). A related framework uses a quantile neural network to model the conditional distribution of the second-stage objective rather than only its expectation, thereby enabling optimization of CVaR-like objectives through an embedded ReLU MIP surrogate (Alcántara et al., 2024). In simulation optimization, a global-local metamodel architecture builds a local Gaussian-process metamodel at each visited first-stage decision to solve many scenario-specific second-stage problems and estimate their optimality gaps, then constructs a global stochastic-kriging metamodel over first-stage decisions that corrects finite-scenario error, second-stage optimality-gap error, and finite exploration error (Xie et al., 2019).

A further coupling mechanism is price-mediated coordination. In interconnected multi-energy systems, the coordinator dualizes the power-balance constraint with multipliers $D$7, producing a local clearing price $D$8. Day-ahead subgradient iterations compute a price vector forecast, while hourly operation updates only the current-period price by bisection (Cheng et al., 2019). In computing power networks, the day-ahead stage solves stochastic unit commitment, while the real-time stage computes economic dispatch and exposes LMP and marginal carbon intensity to a DRL scheduler, which then adapts computational workload placement; the induced CPN load feeds back into the dispatch constraints (Luo et al., 6 Aug 2025).

4. Representative domain interpretations

The phrase “Two-Stage Co-Optimization” is therefore domain-dependent. The following instances illustrate the range of meanings found in the literature.

Domain	Stage 1	Stage 2
High-dimensional BO	Cheap global surrogate with $D$9	Accurate subspace BO with $x$0
Stochastic programming	First-stage decision $x$1	Scenario-dependent recourse $x$2
IMES transactive control	Day-ahead price forecasting	Hourly rolling dispatch
CPN-grid coordination	Day-ahead SUC	Real-time ED plus DRL scheduling
Autonomous driving	Decision-sequence MIP	High-fidelity trajectory planning

In high-dimensional Bayesian optimization, TSCO is an algorithmic coarse-to-fine pattern. CobBO uses a full-space coarse kernel, coordinate-subspace projection, virtual points, and a backoff rule that determines how long to remain in the same subspace before switching (Tan et al., 2021). In stochastic programming, TSCO is the canonical planning/recourse split, extended in some works to revocation and reservation semantics (Zhang et al., 26 Mar 2025, Jiang, 2016).

In energy systems, TSCO becomes a coordination architecture across timescales. The multi-energy-systems framework uses day-ahead market clearing followed by hourly refinement, with privacy and operation authority preserved because MES operators solve their own local problems and exchange only bids and price signals (Cheng et al., 2019). The computing-power-network framework likewise decomposes day-ahead commitment and real-time adaptive scheduling, but adds a carbon-aware DRL agent at the second stage (Luo et al., 6 Aug 2025).

In co-design and autonomy, the notion is more nuanced. The launch-vehicle study criticizes traditional sequential stage-sizing then trajectory-design loops, and its proposed framework jointly optimizes stage masses and trajectory controls in one nonlinear program while still being described as a TSCO framework in the provided summary (Ko et al., 2023). By contrast, YOSO is explicitly presented as a rejection of the conventional two-stage co-design flow; it argues that when the true objective is a composite score involving accuracy, latency, energy efficiency, and QoS, single-stage joint search of the DNN and accelerator can outperform the unidirectional two-stage baseline (Chen et al., 2020). This comparison is important because it shows that TSCO is not always the preferred architecture.

5. Algorithmic guarantees and empirical behavior

The stochastic-programming branch of TSCO has rigorous convergence analyses. The Stochastic Conjugate Subgradient framework establishes a feasible projected direction through null-space projection, proves existence of a step size satisfying sufficient decrease under bounded recourse assumptions, derives a Hoeffding-based sample-size condition scaling like $x$3, and gives an expected stopping-time bound of $x$4-type together with a high-probability near-stationarity guarantee for the projected true subgradient (Zhang et al., 26 Mar 2025). The nonlinear nonconvex decomposition framework proves local differentiability of the barrier-smoothed second-stage solution map, global convergence of the first-stage trust-region SQP for fixed barrier parameter, and superlinear local convergence as the barrier parameter is driven to zero under LICQ, strict complementarity, and second-order sufficiency (Lou et al., 20 Jan 2025). In discrete two-stage optimization via simulation, the global-local metamodel method proves that all first-stage alternatives are revisited infinitely often, second-stage optimality gaps vanish, and the incumbent value converges almost surely to the true optimum under finite discrete decision spaces, boundedness, positive spatial variances, and consistent GP hyperparameter estimation (Xie et al., 2019).

Empirically, TSCO is often justified by computational tractability and improved solution quality under coupling. CobBO reports solutions comparable to or better than other state-of-the-art methods for dimensions ranging from tens to hundreds while reducing both trial complexity and computational costs (Tan et al., 2021). In interconnected MES coordination, 2S-TC achieved almost the same total cost as SG-RTC in the 15-MES case—about $x$5k yuan versus $x$6k yuan—while reducing iteration counts from about $x$7 average and up to $x$8 max to about $x$9–$x$0 average and $x$1 max (Cheng et al., 2019). In launch-vehicle design, simultaneous stage-and-trajectory optimization reduced lift-off mass from $x$2 kg to $x$3 kg in Case I and from $x$4 kg to $x$5 kg in Case II; in Case II runtime fell from $x$6 s in the iterative method to $x$7 s, and in Case III the traditional method diverged after iteration $x$8 while the simultaneous framework converged to a feasible solution with lift-off mass $x$9 kg (Ko et al., 2023).

The same pattern appears in other application areas. YOSO reports $\omega$0 to $\omega$1 energy reduction or $\omega$2 to $\omega$3 latency reduction at the same level of precision relative to a two-stage baseline on the baseline systolic array accelerator and CIFAR-10 (Chen et al., 2020). The hierarchical CPN-grid framework reports lower operational cost and lower carbon emissions than baseline approaches, more than $\omega$4 reduction in RES curtailment, a $\omega$5 job success rate, and $\omega$6 s average tardiness (Luo et al., 6 Aug 2025). The autonomous-driving two-stage optimizer reports $\omega$7 s average execution time, better safety and efficiency indices than FSM+PID and IDP in CARLA, and no collisions or hazardous behavior in the tested CARLA scenarios (Liu et al., 2024). These outcomes do not establish universal superiority, but they do show that stage coupling can materially improve runtime, feasibility, and objective value when the decomposition aligns with the problem structure.

6. Relation to adjacent paradigms, limitations, and open interpretive issues

A persistent source of ambiguity is that “two-stage” can refer either to problem semantics or to algorithm design. In stochastic programming, the stages are semantic: a first-stage decision precedes uncertainty realization and a second-stage decision reacts afterward (Zhang et al., 26 Mar 2025, Alcántara et al., 2024). In CobBO, the stages are algorithmic: a coarse full-space surrogate is followed by local subspace refinement (Tan et al., 2021). In transactive control, the stages are temporal coordination layers: day-ahead and hourly (Cheng et al., 2019). This suggests that TSCO is an umbrella term spanning several non-identical constructions rather than a uniquely standardized formalism.

Another misconception is that TSCO is inherently superior to single-stage co-optimization. The evidence in the surveyed papers is more conditional. YOSO explicitly argues that the dominant two-stage co-design flow can miss the highest composite score because the hardware objective does not feed back into network search, and therefore proposes a single-stage automated framework instead (Chen et al., 2020). Conversely, in autonomous driving, the full integrated problem is acknowledged to be a hard mixed-integer nonlinear program, so a two-stage decomposition is introduced to obtain coherent decisions and trajectories with high computational efficiency (Liu et al., 2024). The appropriate architecture therefore depends on whether decomposition preserves the critical couplings while remaining computationally manageable.

TSCO methods also inherit domain-specific assumptions and limitations. The stochastic conjugate subgradient framework states that its convergence results are restricted to linear equality constraints $\omega$8 in the first stage (Zhang et al., 26 Mar 2025). The nonconvex decomposition framework relies on local branch consistency from warm-started interior-point subproblem solves and on nondegeneracy conditions for its smooth sensitivity theory (Lou et al., 20 Jan 2025). Neural-network-embedded surrogates are tied to MILP-representable architectures such as ReLU networks; the quantile framework additionally highlights quantile crossing as a modeling issue and motivates the IQNN architecture partly to enforce monotonicity by construction (Alcántara et al., 2024). The revocation model introduces a more realistic economic semantics, but also a stricter coupling between first-stage and second-stage variables that complicates LP rounding and decomposition (Jiang, 2016).

Taken together, the literature portrays TSCO as a principled way to retain system-level coupling while reducing dimensionality, scenario burden, or operational latency. Its strongest forms do not merely partition a problem into two chronological segments; they design the interface between stages—through recourse functions, virtual points, dual prices, learned surrogates, or warm-started sensitivities—so that the second stage is both computationally tractable and decision-relevant to the first.