Multi-Objective Co-Design Optimization

• Multi-Objective Co-Design Formulation is a framework that optimizes interconnected subsystems under multiple conflicting objectives and constraints to achieve Pareto-optimal designs.
• It employs methodologies like scalarization, ε-constraint, and evolutionary algorithms to navigate complex design spaces with discrete, continuous, and hierarchical variables.
• The formulation is applied in robotics, synthetic biology, energy systems, and control, enabling efficient trade-off analysis and modular design decomposition.

A multi-objective co-design formulation formalizes the simultaneous design and optimization of multiple interconnected subsystems, subject to a vector of (typically conflicting) objective functions, physical or operational constraints, and often discrete or combinatorial structure. Each design solution represents a compromise between objectives, and the solution set is distilled as the Pareto front of non-dominated designs. Recent literature establishes mathematical frameworks for multi-objective co-design encompassing robotics, synthetic biology, energy systems, communications, and control.

1. Theoretical Foundations and Formal Problem Definitions

At its core, a multi-objective co-design problem is a constrained multi-objective optimization, where the design variables may encapsulate discrete combinatorial choices, continuous tunables, and sometimes hierarchical or recursive subsystem coupling. A canonical formalism is:

$$\min_{x\in\mathcal{X}} \quad F(x) := [f_{1}(x), f_{2}(x), \dots, f_{k}(x)]^T \quad \text{s.t.} \; x\in \mathcal{F}$$

where:

• $x$ denotes the design (and often operational) decision vector;
• $F(x)$ is a vector of $k$ objectives, each mapping $x$ to $\mathbb{R}$;
• $\mathcal{F}$ encodes all equality and inequality constraints arising from subsystem dynamics, feasibility, and architectural coupling.

Pareto optimality is the standard solution concept: a design $x^*$ is Pareto optimal if $\not\exists x'$ such that $f_{i}(x') \leq f_{i}(x^*)$ for all $i$ and strict inequality holds for some $i$.

Frameworks such as the antichain and partial order approach of Monotone Co-Design Problems (MCDPs) rigorously generalize these concepts to heterogeneous and recursive system graphs, enabling fixed-point characterizations and formal modularity (Censi, 2015). Hierarchical decomposition, convex and nonconvex models, integer and categorical variables, and distributed formulations under graph-based agent communication are variously supported in modern treatments (Noori, 2012, Wilhelm et al., 16 May 2025, Xu et al., 2021).

2. Key Principles: Decision Variables, Objectives, and Constraints

Decision variables in multi-objective co-design are problem specific but typically fall into the following categories:

• Architectural/Structural: binary matrices encoding module interconnections (Otero-Muras et al., 2014), graph grammars describing morphologies (Xu et al., 2021), catalog selections (Wilhelm et al., 16 May 2025), or hardware mappings (Das et al., 2022).
• Continuous design tunables: kinetic parameters, gains, sizing parameters, scheduling assignments, or storage capacities (Hollermann et al., 2019, Das et al., 2016, Sharma et al., 14 Jun 2024).
• Operational variables: states, flows, energy transactions, or control trajectories that respond dynamically to the fixed design (Hollermann et al., 2019, Meyur et al., 21 Aug 2024).
• Hierarchical/Scenario coupling: variables active only under certain system architectures (e.g., family design with shared/unshared components) (Bartoli et al., 14 Apr 2025).

Constraints encode system physics, logical validity, compatibility, resource budgets, and dynamic and operational feasibility. Recursive dependencies and feedback loops are explicitly modeled in MCDPs as resource-to-functionality interconnections, and formally resolved via fixed-point iteration on posets (Censi, 2015).

Table: Summary of Principal Modeling Constructs

Variable Type	Example Usage	Reference
Binary structural	Promoter–gene links, hardware mappings	(Otero-Muras et al., 2014, Das et al., 2022)
Continuous tunables	Synthesis rates, LQR weights, sizing	(Das et al., 2016, Hollermann et al., 2019)
Hierarchical	Family design variables, subsystem selection	(Bartoli et al., 14 Apr 2025, Wilhelm et al., 16 May 2025)
Scenario/Stage	Uncertainty, operational adaptation	(Hollermann et al., 2019, Meyur et al., 21 Aug 2024)

3. Solution Methodologies and Pareto Frontier Computation

Canonical solution approaches include:

• Scalarization: Weighted-sum or preference-conditioned objectives, solved repeatedly for sampled weight vectors to approximate the trade-off surface (Censi, 2015, Bartoli et al., 14 Apr 2025, Meyur et al., 21 Aug 2024). Limitation: poor coverage on non-convex Pareto fronts.
• ε-constraint: Fix all but one objective as constraints, sweep the associated bounds, and optimize the remaining objective at each step; recovers nonconvex regions and discrete/branching trade-offs efficiently (Otero-Muras et al., 2014).
• Branch-and-bound / constraint programming (CP): For problems with combinatorial explosion, e.g., robot and hardware co-design, CP with monotone subsystem decomposition efficiently exploits structural monotonicity and prunes dominated configurations (Wilhelm et al., 16 May 2025).
• Surrogate-based Bayesian optimization and mixtures of experts: For expensive black-box objectives with mixed (categorical, continuous, hierarchical) variables, surrogate-assisted adaptive sampling (e.g., SEGOMOE, EHVI, ParEGO) offers substantial efficiency gains, while maintaining Pareto compliance (Bartoli et al., 14 Apr 2025).
• Graph-theoretic and distributed consensus/projection methods: Agent-based distributed optimization for multidisciplinary, multi-objective design ensures convergence to consensus and Pareto-efficient solutions across asynchronously communicating subspaces (Noori, 2012).
• Fixed-point iteration (feedback case): For recursive co-design graphs, the Pareto-optimal resource allocations are characterized as the least fixed point of a monotone operator over antichains (Pareto sets), solved by Kleene's method (Censi, 2015).
• Evolutionary algorithms (e.g., NSGA-II): For hardware mapping and scheduling problems with discrete/categorical variables and multiple objectives, evolutionary approaches with custom crossover/mutation and domain-specific repair accelerate convergence to well-spread multi-dimensional Pareto sets (Das et al., 2022, Das et al., 2016).

Coverage, diversity, and controllability metrics (hypervolume, GD, IGD, PC-entropy, Avg-PCC) are standard for evaluating Pareto set quality (Roy et al., 2023, Bartoli et al., 14 Apr 2025).

4. Applications Across Domains

Synthetic biology (biocircuit design): Simultaneous optimization of regulatory topologies and kinetic parameters to maximize performance (e.g., inducer response/sensitivity) while minimizing cost or complexity, using global MINLP solvers and ε-constraint multi-objective decomposition (Otero-Muras et al., 2014).

Robot co-design: Discrete morphology (graph grammar), control policies, and task-level performance objectives co-optimized, using neural graph heuristics and partial-order structures to generate non-dominated design-control pairs (Xu et al., 2021, Wilhelm et al., 16 May 2025).

Energy systems and storage: Two-stage (design-operation) models where here-and-now capacity investments are coupled to scenario-based operational policies, with Pareto-set selection using ε-indicator and robust, scenario-based co-design (Hollermann et al., 2019, Meyur et al., 21 Aug 2024).

Hardware–software mapping for edge ML: Simultaneous co-optimization of hardware architectural mappings, layer assignment, spatial/temporal scheduling, and energy/latency/area trade-offs, formulated as multi-objective integer programming solved by custom evolutionary search (Das et al., 2022).

Distributed and multidisciplinary systems: Graph-theoretic decomposition across subspace agents, constrained consensus, and subgradient methods guarantee asymptotic convergence to global optima even with only local knowledge (Noori, 2012).

Control and estimation: LQR/FOPID tuning for delayed fractional order systems with multi-objective trade-offs between reference-tracking and control variation, using evolutionary multi-objective algorithms and median solution extraction for tuning-rule synthesis (Das et al., 2016).

Communication/control co-design: Simultaneous scheduling, local computation, and communication policy selection in multi-agent systems, with explicit modeling of delay-accuracy-freshness trade-offs via Age-of-Information and Whittle index policies (Tripathi et al., 2021).

5. Modular Decomposition, Recursion, and Scalability

Decomposition and hierarchy are dominant themes:

• Monotone Subsystem Decomposition: Formalizes conditions under which global Pareto fronts can be constructed by composing subsystem Pareto sets, dramatically scaling multi-objective combinatorial search in catalog-based or modular designs (Wilhelm et al., 16 May 2025).
• Recursive and feedback-interconnected co-design graphs: MCDP framework guarantees that monotonicity and Scott continuity (in the partial order sense) yield the existence and computability of global Pareto sets even in the presence of cyclic dependencies. Practically, this underpins the tractability of large cyber-physical system co-designs (Censi, 2015).

Empirical studies demonstrate sublinear scaling and tractable solution times for designs enumerating up to $10^{25}$ combinations by exploiting structural monotonicity, symmetry, and modular reuse (Wilhelm et al., 16 May 2025).

6. Pareto Front Interpretation and Design Principle Extraction

Multi-objective co-design explicitly illuminates the fundamental trade-offs and enables decision-makers to select among non-dominated options reflecting distinct priorities, complexity budgets, or operating modes.

Common findings:

• Trade-off surfaces often reveal nonconvexity or clusters corresponding to qualitative design principles (e.g., modular architectures, functional adaptation, "diminishing returns" with increased complexity) (Otero-Muras et al., 2014, Wilhelm et al., 16 May 2025).
• Non-intuitive or previously underexplored designs (e.g., hybrid robot morphologies, indirect regulatory topologies) are systematically surfaced only through robust multi-objective exploration.
• Pareto median or knee solutions provide tuning rules or recommended designs, balancing extremal trade-offs (Das et al., 2016, Hollermann et al., 2019).

7. Extensions: Robustness, Uncertainty, and Automation

Robust multi-objective co-design formulations extend the nominal models to accommodate uncertainty:

• Scenario-based and robust formulations: Explicit inclusion of uncertainty sets (e.g., scenarios, polyhedral parameter spaces) and non-anticipativity constraints, with robust Pareto selection minimizing worst-case optimality gaps (Hollermann et al., 2019, Meyur et al., 21 Aug 2024).
• Automation and cloud-scale workflows: Modular, automated architectures (e.g., CAMEO) enable parallel, scalable design space exploration, embedding stochastic programming, scenario generation, cloud-based execution, and workflow monitoring (Meyur et al., 21 Aug 2024).
• Surrogate/hierarchical GFlowNets and goal-conditioning: Learning-based or graph-structured approaches for efficient uniform sampling of Pareto fronts and controllable trade-off navigation, critical in high-dimensional and blackbox-software settings (Roy et al., 2023, Bartoli et al., 14 Apr 2025).

---

The field continues to generalize co-design methodologies for complex, cyber-physical, and data-driven systems, focusing on formal guarantees, efficient scalability, and realistic engineering trade-offs manifested in Pareto-optimal solution spectra (Censi, 2015, Wilhelm et al., 16 May 2025, Das et al., 2022, Otero-Muras et al., 2014, Meyur et al., 21 Aug 2024).