Abstract
Traditional multiobjective design optimization typically focuses on identifying single Pareto-optimal points. However, these solutions can be overly rigid in the face of real-world manufacturing tolerances. This paper introduces a novel methodology using multivariate regression trees to identify optimal design regions (surfaces). By partitioning the design space into high-performance subsurfaces, this approach provides engineers with a range of viable parameters that maintain robustness and performance.
Introduction: Beyond Point Solutions
In engineering design, objectives such as weight, stiffness, and strength often conflict. Traditional optimization techniques aim to find the "best" compromise—a Pareto-optimal point. While mathematically precise, these points offer no flexibility; if a manufacturer cannot meet the exact specification due to tool limitations or cost, the impact on performance is often unknown.
The authors address this by seeking "surface" solutions—regions in the design space where responses are optimal and variances are minimal. This provides designers with the flexibility to adjust parameters within a safe range without compromising the overall integrity of the design.
Methodology: Multivariate Regression Trees
The proposed approach utilizes inductive learning to grow multivariate regression trees. The algorithm recursively partitions the design space into smaller, simpler subsurfaces. The core innovation lies in the splitting rules used to determine the best partitions. The paper evaluates seven distinct criteria:
- Statistical (Non-fuzzy): Including the trace and determinant of the covariance matrix.
- Fuzzy Logic-based: New rules leveraging Pearson’s and Kendall’s measures of association to handle non-linear relationships.
- Decision-Theoretic: Utilizing Bellman and Zadeh’s fuzzy decision-maximizing approach.
- Information-Theoretic: A multidimensional extension of fuzzy entropy.
By incorporating fuzzy logic, the system can better navigate the "vagueness" of design landscapes, discovering complex relationships that standard linear statistics might overlook.
Case Study: Multiobjective Beam Design
The methodology was validated using a classic engineering benchmark: the multiobjective design of a beam structure.
- Design Variables: Length of part 1 ($X_1$) and interior diameter ($X_2$).
- Objectives: Minimize volume ($F_1$), static compliance ($F_2$), and bending stress ($\sigma_g$).
- Data Generation: 900 distinct points were uniformly sampled from the design region, with training sets varying from 100 to 200 points to test model scalability.
Key Results and Discussion
The regression tree approach successfully transitioned the optimization problem from finding points to identifying regions. Leaf nodes in the tree naturally represent clusters of optimal design parameters.
As shown in the results table, while traditional vector optimization (Method "O") provides specific points, the tree-based methods provide ranges. For instance, Method 7 identified a region for $X_1$ between 132.4 and 159.9. Simulations with 5,000 points verified that these regions were consistently high-performing, with predicted volumes (e.g., 3.6) matching verified means exactly and exhibiting very low standard deviations.
Furthermore, the fuzzy splitting criteria (such as those based on Kendall’s $\tau$) matched or exceeded the accuracy of established statistical methods, offering superior flexibility in navigating non-linear design spaces. Because the output is a tree, the logic behind the optimal regions is easily interpretable by human designers.
Conclusion
Inducing multivariate regression trees offers a robust alternative to traditional point-based optimization. By defining optimal regions rather than single points, engineers gain critical flexibility for manufacturing. Future work will focus on scaling this approach to higher-dimensional design spaces and refining pruning strategies to prevent overfitting in complex multivariate scenarios.
Summary updated