Two-Level Decomposition Strategy in Optimization

• Two-Level Decomposition Strategy is a bilevel reformulation that separates non-regular difficulties (e.g., nonconvexity, discontinuity) from high-dimensional, well-behaved parts.
• It employs evolutionary algorithms for irregular upper-level problems and classical solvers for structured lower-level components to enhance robustness and efficiency.
• Numerical validation shows that this method achieves stable, feasible solutions with superior performance compared to traditional single-level approaches.

Searching arXiv for the primary paper and closely related decomposition literature. The two-level decomposition strategy is a bilevel reformulation in which a single-level optimization problem is rewritten so that different kinds of difficulty are handled at different levels. In the formulation developed in "Decomposition of Difficulties in Complex Optimization Problems Using a Bilevel Approach" (Sinha et al., 2024), non-regular parts of a problem—such as nonconvexity, multimodality, discontinuity, non-differentiability, and discrete structure—are assigned to an upper level, while high-dimensional but continuous, differentiable, convex components are assigned to a lower level. The purpose is not to alter the original optimization target, but to separate difficulties and apply different algorithmic paradigms where they are most appropriate (Sinha et al., 2024).

1. Formal reformulation as a bilevel problem

The strategy is presented through the standard bilevel optimization template

$$\begin{aligned} \min_{x \in \mathbb R^n,\,y \in \mathbb R^m}\;& F(x,y) \ \text{subject to}\;\;& G_i(x,y)\le0,\quad i=1,\dots,I,\ & H_j(x,y)=0,\quad j=1,\dots,J,\ & y \in \underset{y}{\arg\min}\;\Big\{\,f(x,y)\;|\; g_k(x,y)\le0,\;\;h_l(x,y)=0,\;\;k=1,\dots,K,\;l=1,\dots,L\Big\}. \end{aligned}$$

Here $x$ are the upper-level variables, $y$ are the lower-level variables, $F$ and $f$ are the upper- and lower-level objectives, $G_i,H_j$ are upper-level constraints, and $g_k,h_l$ are lower-level constraints (Sinha et al., 2024).

The central device is a reformulation of an existing single-level problem. Starting from

$$\min_{x\in\mathbb R^n}\;F(x) \quad \text{s.t.}\quad G_i(x)\le0,\;H_j(x)=0,$$

the variables and constraints are split into two groups according to the type of difficulty they introduce. Writing $x=(x_{k^+},x_{k^-})$, the identical problem can be re-cast as

$$\begin{aligned} \min_{x_{k^+},\,x_{k^-}}\;& F\bigl(x_{k^+},x_{k^-}\bigr)\ \text{s.t.}\quad & G_{i^+}\bigl(x_{k^+},x_{k^-}\bigr)\le0,\quad H_{j^+}\bigl(x_{k^+},x_{k^-}\bigr)=0,\ & x_{k^-}\in\underset{z}{\arg\min}\;\Big\{\,F\bigl(x_{k^+},z\bigr)\;|\; G_{i^-}\bigl(x_{k^+},z\bigr)\le0,\;H_{j^-}\bigl(x_{k^+},z\bigr)=0\Big\}. \end{aligned}$$

This construction preserves the original solution: the same solution $x$0 is recovered (Sinha et al., 2024). A common misunderstanding is that the method changes the optimization problem; in the stated framework, it instead reorganizes the same problem so that its heterogeneous difficulties are exposed structurally.

2. Classification of difficulties and assignment to levels

The decomposition criterion is difficulty type. The paper distinguishes two broad categories. The first is non-regularities, explicitly including nonconvexity, multimodality, discontinuity, non-differentiability, and integer or discrete variables. The second is high dimensionality, meaning large numbers of variables or constraints (Sinha et al., 2024).

The proposed assignment is asymmetric. All non-regular variables and constraints are grouped at the upper level, whereas the high-dimensional but well-behaved part is grouped at the lower level. The rationale is method-specific: evolutionary algorithms are used where continuity, convexity, or differentiability cannot be relied upon, and classical solvers are used where smooth structure can be exploited (Sinha et al., 2024).

Level	Assigned difficulty	Solution approach
Upper level	Nonconvexity, multimodality, discontinuity, non-differentiability, integer or discrete variables	Evolutionary algorithms
Lower level	High-dimensional, continuous, differentiable, convex parts	LP, Interior-Point, or SQP

A concrete example is given for TP1, where the upper level contains $x$1 because they cause a nonconvex objective, while the lower level contains $x$2, described as linear and high-dimensionalizable (Sinha et al., 2024). This suggests that the decomposition is not merely geometric or algebraic partitioning of variables, but an assignment driven by solver compatibility.

3. Algorithmic realization

At the upper level, the method uses a standard Genetic Algorithm. The specified population size is

$x$3

with Simulated Binary Crossover (SBX) at probability $x$4 and Polynomial Mutation at probability $x$5 (Sinha et al., 2024). The upper level also uses the BLEAQ-II framework to gradually approximate the lower-level reaction set mapping $x$6 and optimal-value function $x$7 (Sinha et al., 2024).

At the lower level, the solver depends on structure. If the lower-level problem is linear, a Linear Programming solver is used. If it is smooth nonlinear, Interior-Point or Sequential Quadratic Programming is used. The tolerance on feasibility is set to $x$8 (Sinha et al., 2024).

The interaction between levels is explicit. The GA proposes $x$9, and the classical solver returns

$y$0

The pair $y$1 is then evaluated by $y$2 (Sinha et al., 2024).

The overall framework proceeds by population evolution at the upper level, with each candidate either using a reliable surrogate for $y$3 or triggering a lower-level solve. New lower-level solutions are used to update surrogate models, after which selection, SBX crossover, and polynomial mutation produce the next upper-level population. Termination is defined by best-improvement $y$4 over the last 50 generations (Sinha et al., 2024).

Two operational features are central. First, lower-level solves can be performed in parallel over upper-level individuals. Second, surrogate approximations of the reaction mapping are intended to reduce the number of full lower-level optimizations as the search proceeds (Sinha et al., 2024). A plausible implication is that the decomposition is designed not only for solver specialization but also for amortizing repeated lower-level evaluations.

4. Theoretical interpretation and methodological significance

The theoretical claim is equivalence: by construction the bilevel reformulation is equivalent to the original problem and thus converges to the true optimum under mild assumptions, specifically continuity of $y$5 and existence of a lower-level optimum (Sinha et al., 2024). In addition, surrogate-based approximations of $y$6 are stated to enjoy error-controlled convergence, so overall algorithm convergence improves as surrogate fidelity improves (Sinha et al., 2024).

The practical claims are organized around four properties. The method is described as providing scalability, because the classical solver handles high-dimensional, convex or differentiable blocks efficiently; robustness, because the EA tolerates non-regularities without specialized coding for each new difficulty; computational efficiency, because lower-level surrogates reduce costly lower-level optimizations; and solution quality, because the framework is intended to combine strengths that are separated in purely evolutionary or purely classical approaches (Sinha et al., 2024).

A second common misconception is that the lower level is simply a residual subproblem. The formulation is more specific. The lower level is the part that is high-dimensional yet well behaved, while the upper level retains variables and constraints that induce non-regular behavior. This division is therefore methodological rather than hierarchical in the ordinary master-subproblem sense.

The paper also positions flexibility as a design principle: evolutionary algorithms are useful in this context because their flexibility makes pairing with other solution approaches easy (Sinha et al., 2024). This suggests that the strategy is best understood as a solver-allocation framework embedded in bilevel structure, rather than as a new standalone optimizer.

5. Numerical validation

The numerical study uses ten test problems, TP1–TP10. TP1–TP2 are classic engineering benchmarks, while TP3–TP10 are newly extended and scalable via extra variables $y$7 to induce multimodality, nonseparability, discontinuity, and related effects (Sinha et al., 2024). Comparative methods are grouped as MP, EV, and BO: MP uses Interior-Point or SQP from MATLAB, EV uses the same GA applied directly to the full problem, and BO denotes the bilevel decomposition under BLEAQ-II with GA at the upper level and IP, SQP, or LP at the lower level (Sinha et al., 2024).

The reported setup is MATLAB on Xeon Gold with 32 GB RAM, constraint tolerance $y$8, 11 independent runs per problem, termination by no upper-level improvement in 50 generations for BOBD, the same wall-time budget for GA, and default settings for MP (Sinha et al., 2024).

The empirical findings are specific. BOBD produced feasible solutions in every run. MP and EAs often returned infeasible or poor-quality points when confronted with nonconvexity or high dimensionality (Sinha et al., 2024). On TP1, BOBD found

$y$9

at

$F$0

improving upon the best-known $F$1 (Sinha et al., 2024). Across TP1–TP10, BOBD’s best, median, and worst solutions coincided, which the paper interprets as stability (Sinha et al., 2024). In two additional test beds obtained by adding 20 and 50 scalable variables $F$2, BOBD continued to dominate (Sinha et al., 2024).

The comparative pattern is also explicit. MP is described as good on convex-like cases but often infeasible or suboptimal on multimodal or nonconvex cases. GA alone handles nonconvexity but degrades rapidly as dimension or constraints increase. BOBD is presented as integrating both strengths (Sinha et al., 2024). The numerical evidence therefore supports the intended specialization principle: non-regularities are handled by the EA, and structured scale is handled by classical mathematical programming.

6. Relation to other two-level decompositions

The phrase two-level decomposition is used in multiple technical literatures, but not always with the same meaning. In the BOBD setting, the decomposition is over types of difficulty and solver roles (Sinha et al., 2024). In other arXiv works, the two-level structure has a different semantics.

In "Sequential Relational Decomposition" (Fried et al., 2019), a task relation $F$3 is factored into sequential subtasks $F$4 and $F$5, with total and partial decomposition decision problems. There the central issue is representational factorization and its complexity: explicit TDP/PDP is NP-complete, symbolic TDP/PDP is NEXPTIME-complete, and automatic TDP/PDP is conjectured undecidable (Fried et al., 2019). This is a decomposition of behavior into stages, not a decomposition of optimization difficulty.

In "A Decomposition Framework for Nonlinear Nonconvex Two-Stage Optimization" (Lou et al., 20 Jan 2025), the structure is again different. A first-stage master problem in $F$6 is coupled to second-stage value functions $F$7, and smoothing via an interior-point formulation makes the second-stage solution locally differentiable with respect to first-stage parameters. The resulting method is a two-level algorithm with a smoothed master and warm-started interior-point subproblems (Lou et al., 20 Jan 2025). Here the decomposition is stagewise and scenario-based.

In computational PDE and scientific computing, two-level often denotes a fine/coarse hierarchy. The GDSW preconditioner combines local overlapping Schwarz solves with a global coarse solve in an energy-minimizing subspace (Yamazaki et al., 2023). The two-level Deep-DDM adds a global coarse network $F$8 to fine subdomain PINNs and restores numerical scalability by transporting low-frequency or global information in $F$9 iterations (Dolean et al., 2024). These are decompositions across resolution levels.

Taken together, these usages indicate that two-level decomposition is a family resemblance term rather than a single fixed concept. A plausible implication is that BOBD is distinctive because its split is neither temporal, spatial, nor purely algebraic; it is a decomposition of heterogeneous mathematical difficulties into levels chosen to match complementary optimization paradigms.