Decoupling Constraints from Two Directions for Evolutionary Constrained Multi-objective Optimization

Technical Summary: Decoupling Constraints from Two Directions for Evolutionary Constrained Multi-objective Optimization

Problem Statement
Real-world constrained multi-objective optimization problems (CMOPs) often involve multiple constraints with complex coupling relationships. Traditional constraint-handling techniques (CHTs) typically aggregate all constraint violations into a single scalar value, ignoring the distinct roles and interactions of individual constraints. While recent constraint-decoupling methods address this by handling constraints separately, they generally search only in the "evolutionary direction" (toward nondominated regions) to approximate single-constraint Pareto fronts (SCPFs).

The authors identify a critical limitation in existing approaches: a portion of the constrained Pareto front (CPF) may be unrelated to any SCPF. Instead, this portion, termed the Independent CPF (ICPF), is shaped by the boundaries of infeasible regions. In such cases, searching solely in the evolutionary direction fails because the optimal solutions are not dominated by any solution satisfying a single constraint; rather, they emerge from the joint effect of multiple constraints. Consequently, existing methods may miss the ICPF or fail to navigate the infeasible barriers that define it.

Methodology: DCF2D
To address these challenges, the paper proposes DCF2D (Decoupling Constraints from Two Directions), a bidirectional constraint-decoupling coevolutionary algorithm. The core innovation lies in the introduction of the Reverse CPF (RCPF), which characterizes informative infeasible boundaries that dominate the ICPF. DCF2D dynamically identifies constraints obstructing the main population and activates constraint-specific auxiliary populations that search in either the evolutionary direction (for SCPFs) or the reverse direction (for RCPFs).

The algorithm operates through a three-stage framework:

1. Global Exploration: An unconstrained auxiliary population ($AP_0$) explores the unconstrained Pareto front (UPF) to establish a baseline. A sliding-window trajectory analysis monitors convergence to determine when to transition to the next stage.
2. Event-Driven Bidirectional Coevolution: This stage dynamically manages constraint decoupling.
   • Activation: An infeasible archive ($LI$) stores informative infeasible solutions that dominate the main population. When $LI$ identifies constraints currently blocking convergence, specific auxiliary populations ($AP_i$) are activated.
   • Direction Switching: Each active $AP_i$ initially searches in the evolutionary direction to approximate the SCPF. If no feasible solution is found (indicating the CPF is disjoint from the SCPF), the search direction switches to the reverse direction to approximate the RCPF. Conversely, if an RCPF search population becomes dominated by the main population, it is deactivated.
   • Resource Allocation: Computational resources are shared between the main population and active auxiliary populations. Constraint-specific populations are smaller ($N/4$) to maintain efficiency while tracking boundaries.
3. Final Convergence: Once a predefined fraction of the evaluation budget ($\beta \times MaxFE$) is reached, all auxiliary populations are deactivated. The remaining resources are dedicated to the main population to refine the convergence and distribution of the final feasible solutions.

Key mechanisms include a two-step update for the infeasible archive ($LI$) to capture boundary solutions and a hierarchical selection strategy for auxiliary populations that balances objective values with constraint violations to prevent drift.

Key Contributions

1. Theoretical Analysis: The authors analyze how multiple constraints jointly shape the CPF, formally defining the ICPF and RCPF. This provides a geometric interpretation of constraint coupling, demonstrating that parts of the CPF can only be discovered by searching in the reverse evolutionary direction from infeasible boundaries.
2. Dynamic Mechanism: A novel event-driven mechanism is developed to identify active constraints and determine the optimal search direction (evolutionary vs. reverse) for each constraint. This avoids unnecessary computational overhead associated with decoupling irrelevant constraints.
3. Algorithm Proposal: The DCF2D algorithm is proposed to exploit constraint coupling in both directions. It integrates unconstrained global exploration, adaptive bidirectional coevolution, and final convergence refinement.

Experimental Results
The performance of DCF2D was evaluated on 87 benchmark instances from seven test suites (DASCMOP, DOC, LIRCMOP, MW, ZXH_CF, CDTLZ, SDC) and 28 real-world engineering CMOPs (RWMOPs). It was compared against nine state-of-the-art algorithms, including five constraint-decoupling methods and four high-performance CMOEAs.

• Overall Performance: DCF2D achieved the best overall average rank (2.73) across all test suites, significantly outperforming the second-ranked algorithm (IMTCMO, 3.87) based on the Nemenyi post-hoc test.
• Benchmark Suites: DCF2D demonstrated superior performance on suites with complex constraint interactions, particularly DASCMOP, DOC, and CDTLZ. It showed particular strength on LIRCMOP, where large infeasible regions obstruct evolutionary search, by successfully utilizing reverse-direction guidance.
• Real-World Applications: On the 28 RWMOPs, DCF2D achieved the best average rank (3.46) in terms of Hypervolume (HV), outperforming both aggregation-based and other decoupling methods.
• Scalability: The algorithm's performance improved relative to competitors as the number of constraints increased, achieving the best rank when the number of constraints was 14.
• Ablation Study: Variants of DCF2D using only evolutionary search (DCF2D_EVO) or only reverse search (DCF2D_REV) performed significantly worse, confirming the necessity of bidirectional guidance.

Significance and Claims
The paper claims that effective constraint decoupling must not only identify relevant constraints but also determine the appropriate search direction based on the geometric relationship between the CPF and infeasible regions. By introducing the RCPF and the bidirectional search strategy, DCF2D addresses the limitation of existing methods that fail when the CPF is bounded by infeasible regions rather than SCPFs.

The authors conclude that DCF2D is particularly effective for problems with disconnected feasible regions or extensive infeasible barriers. However, they modestly acknowledge limitations: the algorithm decouples constraints individually, potentially overlooking interactions among strongly coupled composite constraints, and the Pareto-dominance-based screening of the infeasible archive may lose selection pressure in many-objective scenarios. Future work is suggested to explore automatic constraint grouping and indicator-based archive management.