TL;DR
This paper introduces TC-SAEA, a novel evolutionary algorithm designed for multi-objective optimization where one objective is significantly slower to evaluate than the other. By using a "co-surrogate" to model the relationship between fast and slow objectives and a smart transfer learning filter, the algorithm effectively "borrows" knowledge from cheap evaluations to accelerate the search for expensive ones.
What Problem Does This Paper Address?
In the real world, optimization isn't always symmetrical. Imagine designing an aircraft: calculating its aerodynamic properties might take a few hours (the "fast" objective), but simulating a full-scale crash test could take hundreds of hours (the "slow" objective).
Standard Multi-Objective Evolutionary Algorithms (MOEAs) usually assume all objectives can be evaluated simultaneously. When evaluation times are non-uniform, these algorithms often sit idle, waiting for the slow objective to finish. This leads to a massive waste of computational resources and slow convergence. While some researchers have tried "fitness inheritance" or simple interleaving, they often struggle to maintain accuracy or fail to exploit the underlying functional relationship between the different objectives.
Methods
The authors propose the Transfer Learning Based Co-surrogate Assisted Evolutionary Algorithm (TC-SAEA). Its architecture rests on three pillars:
- The Co-surrogate Model ($GP_c$): Instead of just modeling the slow objective directly, TC-SAEA builds a Gaussian Process (GP) model to predict the difference between the slow and fast objectives ($f_s - f_f$). Because objectives near the Pareto front are often highly correlated, this "delta" is often easier to learn than the raw function.
- Synthetic Data Generation: While waiting for the slow objective to finish, the algorithm performs extra evaluations on the fast objective. It then uses the co-surrogate to "guess" the corresponding values for the slow objective, creating a set of synthetic training instances.
- Transferable Instance Selection: Not all guesses are good. To prevent "negative transfer" (where bad data ruins the model), the algorithm uses the uncertainty (standard deviation) provided by the slow objective's primary surrogate ($GP_s$). It only accepts synthetic data points that fall within a specific confidence interval ($CI = Y_s^a \pm \sigma_s^a$).
- Model Management: The algorithm uses an adaptive acquisition function and Angle-Penalized Distance (APD) to select which new solutions are worth the "expensive" evaluation, ensuring a balance between exploring new areas and refining the current best solutions.
Datasets
The researchers evaluated TC-SAEA using several standard and modified benchmarks:
- DTLZ Suite (DTLZ1-DTLZ7): Classic multi-objective problems, including modified versions (DTLZ1a, DTLZ3a) with reduced ruggedness to test convergence capabilities.
- UF Suite (UF1-UF7): Problems with complex Pareto set shapes.
- cm-OneMax: A continuous version of the OneMax problem used specifically to test how the algorithm handles different levels of correlation (positive, negative, or none) between objectives.
Key Insights and Ideas
- The Power of the "Delta": The "aha moment" here is that the relationship between objectives is often more stable than the objectives themselves. By modeling $f_s - f_f$, the co-surrogate captures the "physics" of the trade-off.
- Uncertainty as a Gatekeeper: Transfer learning in optimization is risky; if you feed a surrogate model wrong data, the evolutionary search can go off the rails. Using the GP's own uncertainty as a filter for synthetic data is a clever way to ensure "positive transfer."
- Exploiting Idle Time: The algorithm treats the "waiting period" for slow evaluations not as a bottleneck, but as an opportunity to gather more data on the cheap objective, which is then transformed into useful knowledge for the expensive one.
Key Results
The authors compared TC-SAEA against several state-of-the-art methods (like HK-RVEA and T-SAEA) across different time-ratio settings ($\tau = 5$ and $\tau = 10$).
- Superior Convergence: TC-SAEA outperformed the competition on the majority of DTLZ and UF test instances, achieving lower Inverted Generational Distance (IGD) values, which indicates a closer and more diverse approximation of the true Pareto front.
- Robustness to Correlation: On the cm-OneMax problem, TC-SAEA showed the best performance regardless of whether the objectives were positively or negatively correlated, proving the versatility of the co-surrogate approach.
- Efficiency: The algorithm demonstrated much 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).
Limitations and Future Work
- Biased Pareto Fronts: The algorithm struggled slightly on DTLZ4. The authors note that when the density of points on the Pareto front is highly biased, the synthetic values can be pushed too far from the truth, limiting the benefits of transfer.
- Scaling to Many Objectives: Currently, the work focuses on bi-objective problems. Extending this to "many-objective" optimization (3+ objectives) where multiple objectives might have different latencies remains a challenge.
- Resource Allocation: The paper assumes a fixed integer ratio ($\tau$) for evaluation times. Future work could explore dynamic environments where evaluation times vary or where the algorithm must decide how to optimally allocate a total time budget between different objectives.
Summary updated