Abstract
Large-scale multiobjective optimization (LSMOP) is essential for solving complex engineering problems with hundreds of variables and conflicting goals. However, existing evolutionary algorithms often suffer from high performance sensitivity, yielding inconsistent results across different runs due to stochastic initialization. This paper proposes LMOMCTS, a framework that reframes the evolutionary process as a Monte Carlo Tree Search (MCTS). By leveraging tree-based exploration and decision variable sampling, LMOMCTS achieves high-quality, reliable solutions. Evaluations on standard benchmarks and real-world industrial datasets show that LMOMCTS significantly improves performance insensitivity while maintaining state-of-the-art convergence and diversity.
Introduction
Large-scale multiobjective optimization (LSMOP) involves optimizing hundreds or thousands of decision variables across multiple conflicting objectives. While Evolutionary Algorithms (EAs) are commonly used, they face a significant "stochasticity problem." Random initialization and probabilistic operators can lead to high variance in performance, where one run succeeds and another fails.
In industrial settings, where computational resources are costly, engineers require algorithms that are "insensitive" to such randomness—providing stable, high-quality results consistently. This research addresses this gap by focusing on "performance insensitivity," ensuring that an algorithm's performance remains robust regardless of the initial random seed.
Methodology: The LMOMCTS Framework
The proposed framework, LMOMCTS (Large-scale Multiobjective Optimization via Monte Carlo Tree Search), treats the optimization process as a tree search problem. The core innovation lies in its four-step approach:
- MCTS Framework: It utilizes the Upper Confidence Bound applied to Trees (UCT) to navigate the search space. Each node in the tree represents a population of solutions, allowing the algorithm to maintain multiple potential evolutionary directions.
- Decision Variable Sampling for Optimization (DVSO): To combat the "curse of dimensionality," the algorithm samples subsets of decision variables during node expansion. This creates diverse child populations that explore different dimensions of the problem space without the need for complex variable linkage analysis.
- Hypervolume (HV) Evaluation: Promising branches are identified by estimating the Hypervolume of the population—a metric that measures both convergence (closeness to the ideal) and diversity (spread of solutions).
- Backpropagation: Evaluation results update UCT values throughout the tree, guiding the search toward promising regions while maintaining enough exploration to avoid local optima.
Experimental Evaluation
The researchers compared LMOMCTS against eight state-of-the-art algorithms, including CCGDE3 and LMOCSO, across the LSMOP and WFG benchmarks, as well as a real-world voltage transformer configuration dataset (TREE).
Key findings include:
- Superior Consistency: LMOMCTS achieved the best "insensitive-IGD" results in 34 out of 54 test instances, demonstrating that its "worst" runs remain close to its "best" runs.
- General Performance: It outperformed competitors in 30 out of 54 cases for overall convergence and diversity.
- Real-World Application: On the TREE dataset, LMOMCTS secured the best mean results in 5 out of 6 instances, proving its utility in practical industrial scenarios.
- Efficiency: Despite the complexity of maintaining a tree structure, the algorithm maintains stable memory usage and relatively low computation time.
Conclusion
LMOMCTS offers a robust solution for large-scale optimization by prioritizing performance reliability. By maintaining multiple potential evolutionary directions in a tree structure, it avoids the pitfalls of single-path searches that often get stuck in local optima. While it excels in most scenarios, it faces challenges with disconnected Pareto Fronts, suggesting future work in specialized operators or machine learning-guided variable sampling to further enhance its robustness.
Summary updated