Abstract
This paper presents a framework for solving Uncertain Multi-Objective Programming (UMOP) problems characterized by correlated objective functions. By integrating the Expected-Variance ($C_{EV}$) criterion, the authors provide a method to transform uncertain multi-objective models into deterministic single-objective problems. The validity of the approach is proven through Pareto efficiency theorems and demonstrated via continuous and discrete numerical benchmarks. Results indicate that the $C_{EV}$ criterion effectively captures volatility, leading to more robust decision-making than traditional expected-value models.
Introduction
In real-world optimization scenarios—such as power plant scheduling or air traffic management—parameters are rarely fixed. Researchers often utilize "Uncertainty Theory" to model expert belief degrees when data is scarce. However, traditional Uncertain Multi-Objective Programming (UMOP) methods frequently suffer from two limitations: they assume objective functions are independent, and they focus solely on the expected value, ignoring the volatility (variance) of uncertain variables. This paper addresses these gaps by proposing a $C_{EV}$ criterion-based approach that accounts for both correlation and risk.
Proposed $C_{EV}$ Framework
The researchers propose a two-step transformation to convert complex, uncertain problems into solvable, deterministic equivalents:
- Objective Reduction: The multi-objective problem is first converted into an Uncertain Single-Objective Programming (USOP) problem using techniques like the Linear Weighted Method (LWM) or the Ideal Point Method (IPM).
- Deterministic Transformation: Using the $C_{EV}$ criterion, the uncertain objective is transformed into a deterministic function that combines the Expected Value and the Variance.
Mathematical proofs provided in the study confirm that the optimal solution of this deterministic model remains a "Pareto efficient" solution for the original uncertain problem, ensuring that no objective can be improved without degrading another.
Numerical Experiments
The approach was evaluated using two sophisticated benchmarks. The first is a continuous optimization problem involving three correlated objective functions and two uncertain variables ($\xi_1, \xi_2$). The setup for these uncertain variables is detailed in the table below.
The second benchmark is a discrete 50-item "backpack problem" (knapsack problem) with uncertain weights and values modeled using log-normal distributions. To solve these models, the authors employed metaheuristic algorithms: Genetic Particle Swarm Optimization (G-PSO) for continuous cases and a Binary Wolf Pack Algorithm for discrete cases.
Results and Discussion
The $C_{EV}$ approach was compared against traditional expected-value methods. In the continuous case (Example 1), the proposed method yielded distinct optimal solutions $(0.5748, 2.4739)$, whereas the traditional method resulted in $(0.5493, 2.6815)$. The $C_{EV}$ method provided a more comprehensive view of the trade-offs by maintaining the uncertainty profile of the problem.
In the discrete backpack problem, the $C_{EV}$ approach achieved a more balanced result across three objectives while strictly adhering to resource constraints. The algorithms demonstrated strong convergence, with the Wolf Pack algorithm stabilizing by the 370th generation.
Conclusion
By incorporating variance into the objective function, the $C_{EV}$ model inherently favors more stable, less risky solutions. This framework bridges the gap between abstract Uncertainty Theory and practical engineering by handling correlated objectives without losing the essence of multi-objective trade-offs. Future work will explore extending this approach to multi-level programming problems involving multiple hierarchical decision-makers.
Summary updated