Multi-Objective Geometric Programming

• Multi-Objective Geometric Programming is a framework for optimizing multiple posynomial objectives simultaneously, balancing trade-offs in engineering and finance.
• It employs aggregation methods like the weighted-sum approach to generate Pareto frontiers, enabling detailed trade-off analysis.
• Advanced numerical techniques, including duality theory and sequential convex programming, address challenges like nonconvexity and high degrees of difficulty.

Multi-objective geometric programming (MOGP) is the study and solution of geometric programming problems in which multiple posynomial objective functions must be optimized simultaneously, often subject to shared posynomial constraints. This domain is central to large-scale engineering design, financial management, and operations research, where practitioners are typically required to evaluate trade-offs between several objectives that may conflict. Classical geometric programming (GP) provides a framework for single-objective optimization of functions with posynomial structure, but MOGP requires both new theoretical developments and specialized numerical methodologies to produce a representative set of non-inferior (Pareto-optimal) solutions, exposing trade-offs to decision-makers (Ojha et al., 2010, Ojha et al., 2010, 0912.1832).

1. Problem Formulation

A general MOGP is defined over $n$ strictly positive real variables $x = (x_1, \dots, x_n)$ with $p$ posynomial objectives and $m$ posynomial inequality constraints:

$$\begin{aligned} &\min_{x > 0} \quad F(x) = (f_1(x), f_2(x), \ldots, f_p(x)) \ &\text{subject to:} \quad g_i(x) = \sum_{t=1}^{T_i} C_{it} \prod_{j=1}^n x_j^{a_{itj}} \leq 1, \quad i=1, \ldots, m \end{aligned}$$

where each $f_k(x)$ is a posynomial:

$$f_k(x) = \sum_{t=1}^{T_{k0}} g_k(t) C_{k0t} \prod_{j=1}^n x_j^{a_{k0tj}}$$

with $C_{k0t}, C_{it} > 0$, $a_{k0tj}, a_{itj} \in \mathbb{R}$, and $g_k(t)$ a continuous positive mapping over a parameter $x = (x_1, \dots, x_n)$0 (Ojha et al., 2010). When cost coefficients are parameterized as continuous functions, MOGP models a wider range of engineering design scenarios.

2. Weighted-Sum Method and Pareto Front Generation

The standard approach for scalarizing MOGP is the weighted-sum (weighted mean) method: introduce a weight vector $x = (x_1, \dots, x_n)$1 with $x = (x_1, \dots, x_n)$2, and form the aggregate single-objective GP:

$x = (x_1, \dots, x_n)$3

The resulting aggregate $x = (x_1, \dots, x_n)$4 retains posynomial structure and is thus solvable by geometric programming duality. By sweeping $x = (x_1, \dots, x_n)$5 over the probability simplex, one traces the trade-off surface between objectives, i.e., the (convex) Pareto frontier (Ojha et al., 2010, Ojha et al., 2010). All objectives must be in minimization posynomial form; maximizations are recast through inversion or sign changes in exponents (Ojha et al., 2010).

If the multi-objective attainable set is convex, every Pareto-optimal point is reachable for some $x = (x_1, \dots, x_n)$6. Non-convex portions of the Pareto set cannot be obtained by the weighted-sum strategy alone.

3. Duality Theory and Solution Algorithms

After aggregation, the single-objective posynomial GP is solved by duality:

• Dual variables $x = (x_1, \dots, x_n)$7, one per objective-term, and $x = (x_1, \dots, x_n)$8, one per constraint-term, are introduced.
• The dual objective is:

$x = (x_1, \dots, x_n)$9

subject to:

$p$0

$p$1

where the $p$2 collect the weighted, parameterized terms of all objectives. Optimality is guaranteed via Karush-Kuhn-Tucker (KKT) conditions. Primal optimal variables $p$3 can be recovered directly from the dual solution.

The algorithmic flow consists of iterating the above procedure for a grid of weights $p$4, recovering the corresponding Pareto set (Ojha et al., 2010).

4. Alternative Methods: Lexicographic Order and Extensions

Lexicographic multi-objective geometric programming (LGP) optimizes objectives by strict priority. Objectives are ordered (ranked), and one solves a succession of GPs:

1. Minimize $p$5 under original constraints.
2. Minimize $p$6 subject to original constraints plus $p$7.
3. Continue sequentially, adding upper bounds for all previously minimized objectives (0912.1832).

Each subproblem is a standard GP, with duality theory and strong duality applying as in scalar GP. This yields unique, lex-minimal Pareto solutions when exponent-matrix rank conditions are satisfied.

Signomial extensions, nonconvex objectives, and applications to parametric Markov decision processes (MDPs) have motivated sequential convex programming approaches: signomial programs are convexified locally as a sequence of GPs, solved iteratively in a trust-region framework. Each GP enforces feasibility with respect to the original specifications, and scalarization is applied to the multi-objective cost vector as above (Cubuktepe et al., 2017).

5. Computational Aspects and Degree of Difficulty

The computational tractability of MOGP scalarized using the weighted-sum method is directly linked to the so-called “degree of difficulty” (DOD), defined as:

$p$8

• DOD = 0: unique dual solution.
• DOD > 0: dual is a concave maximization over a polyhedron (solvable by standard methods).
• DOD < 0: possible infeasibility; global nonlinear optimization may be required (Ojha et al., 2010).

All dual problems resulting from posynomial objectives and constraints are convex, linearly-constrained, and efficiently solvable (e.g., using interior-point methods). Nonconvexity in the aggregate objective space limits coverage of the Pareto frontier by the weighted-sum approach.

6. Representative Numerical Examples

Published works illustrate the key techniques with explicit multi-objective scenarios. For instance, Ojha & Das (Ojha et al., 2010) analyze a MOGP with parameterized cost coefficients $p$9, $m$0 in three variables, generating Pareto solutions at fixed weights $m$1:

Weight $m$2	Optimal $m$3	$m$4
(0.1, 0.9)	$m$5	51.4067
(0.5, 0.5)	$m$6	75.8160

By enumerating $m$7 across the simplex, a table of Pareto points is produced, enabling detailed trade-off analysis.

7. Practical Impact and Limitations

Multi-objective geometric programming is widely utilized in engineering design, resource allocation, and system optimization where trade-off analysis is critical. The weighted-sum method is favored for its algorithmic simplicity and clear interpretation of the relative importance of criteria. Lexicographic MOGP is preferred when objectives admit a strict natural priority; sequential convex programming is effective for high-dimensional or signomial (nonconvex) regimes. Limitations include inability of the weighted-sum to produce non-convex Pareto points and the computational challenge posed by large DOD or nonconvexity (Ojha et al., 2010, Ojha et al., 2010, Cubuktepe et al., 2017, 0912.1832).

Extensions to signomial programming, and applications to the verification and synthesis of parametric MDPs, demonstrate the adaptability of MOGP frameworks for a vast array of engineering and systems applications, maintaining polynomial-time solvability for convex instances and scalability well beyond conventional SAT/SMT-based methods (Cubuktepe et al., 2017).