LiBOG: Lifelong Learning for Black-Box Optimizer Generation

TL;DR

LiBOG addresses the practical challenge of evolving problem distributions in MetaBBO by formulating lifelong learning as a non‑stationary MDP and learning symbolically structured updating rules with PPO. It introduces inter‑task consolidation via elastic weight consolidation and intra‑task consolidation via elite behavior consolidation to mitigate forgetting while preserving plasticity. Empirical results on the IEEE CEC benchmark show LiBOG achieves superior optimization performance and significantly reduces forgetting across multiple task orders, with ablations confirming the value of both consolidation mechanisms. The approach advances automated optimizer design for real‑world, shifting problem landscapes and offers a scalable framework for continual BBO solver generation.

Abstract

Meta-Black-Box Optimization (MetaBBO) garners attention due to its success in automating the configuration and generation of black-box optimizers, significantly reducing the human effort required for optimizer design and discovering optimizers with higher performance than classic human-designed optimizers. However, existing MetaBBO methods conduct one-off training under the assumption that a stationary problem distribution with extensive and representative training problem samples is pre-available. This assumption is often impractical in real-world scenarios, where diverse problems following shifting distribution continually arise. Consequently, there is a pressing need for methods that can continuously learn from new problems encountered on-the-fly and progressively enhance their capabilities. In this work, we explore a novel paradigm of lifelong learning in MetaBBO and introduce LiBOG, a novel approach designed to learn from sequentially encountered problems and generate high-performance optimizers for Black-Box Optimization (BBO). LiBOG consolidates knowledge both across tasks and within tasks to mitigate catastrophic forgetting. Extensive experiments demonstrate LiBOG's effectiveness in learning to generate high-performance optimizers in a lifelong learning manner, addressing catastrophic forgetting while maintaining plasticity to learn new tasks.

Figures (9)

• Figure 1: Existing MetaBBO methods train the BBO optimizer $O$ with data $D_0$ obtained from the problem distribution $P_0$ available during the training phase, then freeze the model after training to solve the newly arising problems from distributions $(P_1,P_2,\dots)$. New data obtained from subsequent problems are discarded. In contrast, lifelong learning MetaBBO utilizes the data obtained from each encountered problem to update the optimizer continually.
• Figure 2: Illustration of LiBOG, with different problem distributions $P_{i-1},P_{i},\dots$ sequentially arrive.
• Figure 3: Test performance on each learned task during the lifelong learning process of task order 0. In fine-tuning, catastrophic forgetting of previous tasks is significant, but is mild in LiBOG.
• Figure 4: Test performance on each task of models obtained during the lifelong learning process, under task order 0. Vertical gray lines indicate the time of task changes.
• Figure 5: Performance of LiBOG trained with different values of $\alpha$ and $\beta$. Under all settings, LiBOG outperforms baselines.