Abstract
Most existing multi-objective evolutionary algorithms (MOEAs) implicitly assume that each objective function can be evaluated within the same period of time. Typically, this is untenable in many real-world optimization scenarios where evaluation of different objectives involves different computer simulations or physical experiments with distinct time complexity. To address this issue, a transfer learning scheme based on surrogate-assisted evolutionary algorithms (SAEAs) is proposed, in which a co-surrogate is adopted to model the functional relationship between the fast and slow objective functions and a transferable instance selection method is introduced to acquire useful knowledge from the search process of the fast objective. Our experimental results on DTLZ and UF test suites demonstrate that the proposed algorithm is competitive for solving bi-objective optimization where objectives have non-uniform evaluation times.
Introduction
In many real-world multi-objective optimization scenarios, evaluation times for different objectives are non-uniform. For instance, in aircraft design, aerodynamic calculations might be significantly faster than full-scale crash simulations. Standard Multi-Objective Evolutionary Algorithms (MOEAs) typically assume uniform evaluation times, leading to computational inefficiencies when one objective is much slower than another. This paper addresses this imbalance by proposing a Transfer Learning Based Co-surrogate Assisted Evolutionary Algorithm (TC-SAEA).
Methodology
The TC-SAEA framework leverages the idle time during slow objective evaluations to perform additional fast objective evaluations. These are then used to augment the training data for the slow objective via a transfer learning scheme.
Co-surrogate Model ($GP_c$)
Instead of modeling the slow objective ($f_s$) directly, TC-SAEA builds a Gaussian Process (GP) model to predict the difference between the slow and fast objectives ($f_s - f_f$). This "delta" is often more stable and easier to learn than the raw function, especially near the Pareto front where objectives are highly correlated.
Transfer Learning and Instance Selection
The algorithm generates synthetic data for the slow objective using the co-surrogate. To avoid "negative transfer," where inaccurate synthetic data degrades model performance, a Transferable Instance Selection method is employed. It uses the uncertainty (standard deviation $\sigma_s^a$) from the primary slow-objective surrogate ($GP_s$) to filter synthetic points. Only instances falling within a defined confidence interval ($CI = Y_s^a \pm \sigma_s^a$) are accepted into the training set $D_t$.
Model Management
The algorithm uses an adaptive acquisition function and Angle-Penalized Distance (APD) to select new solutions for expensive evaluations, ensuring a balance between exploring new areas and refining the current best solutions.
Experimental Evaluation
The performance of TC-SAEA was evaluated on standard benchmarks including the DTLZ and UF suites, as well as the cm-OneMax problem for correlation testing.
Key Results
- Superior Convergence: TC-SAEA outperformed state-of-the-art methods like HK-RVEA and T-SAEA on the majority of DTLZ and UF test instances, achieving lower Inverted Generational Distance (IGD) values.
- Robustness to Correlation: On the cm-OneMax problem, TC-SAEA showed consistent performance regardless of whether the objectives were positively or negatively correlated, proving the versatility of the co-surrogate approach.
- Efficiency: The algorithm demonstrated faster convergence in the early stages of optimization, which is critical when the total budget of expensive evaluations is very limited (e.g., only 200 evaluations).
Conclusion
The TC-SAEA algorithm effectively turns the "waiting time" for slow evaluations into a productive data-gathering phase by capturing the functional relationship between objectives through a co-surrogate and using uncertainty-based filtering for transfer learning. While it shows promise, future work will address limitations such as performance on highly biased Pareto fronts and extending the framework to many-objective optimization with multiple non-uniform evaluation latencies.
Summary updated