DesignX: Human-Competitive Algorithm Designer for Black-Box Optimization

TL;DR

DesignX tackles the challenge of automated black-box optimizer design by introducing an end-to-end framework that jointly learns optimizer workflow generation and dynamic hyperparameter configuration. It builds Modular-EC, a rich modular space of hundreds of components, and trains a cooperative dual-agent Transformer system on a large synthetic distribution of problems ($n=12{,}800$), achieving human-competitive performance on both synthetic benchmarks and realistic tasks. The approach reveals non-trivial design patterns and provides practical insights into which module types matter under different problem properties, while delivering competitive runtime and scalability. Overall, DesignX demonstrates a scalable path toward automatic, problem-specific optimizer design and offers a foundation for further end-to-end automation of algorithm design.

Abstract

Designing effective black-box optimizers is hampered by limited problem-specific knowledge and manual control that spans months for almost every detail. In this paper, we present \textit{DesignX}, the first automated algorithm design framework that generates an effective optimizer specific to a given black-box optimization problem within seconds. Rooted in the first principles, we identify two key sub-tasks: 1) algorithm structure generation and 2) hyperparameter control. To enable systematic construction, a comprehensive modular algorithmic space is first built, embracing hundreds of algorithm components collected from decades of research. We then introduce a dual-agent reinforcement learning system that collaborates on structural and parametric design through a novel cooperative training objective, enabling large-scale meta-training across 10k diverse instances. Remarkably, through days of autonomous learning, the DesignX-generated optimizers continuously surpass human-crafted optimizers by orders of magnitude, either on synthetic testbed or on realistic optimization scenarios such as Protein-docking, AutoML and UAV path planning. Further in-depth analysis reveals DesignX's capability to discover non-trivial algorithm patterns beyond expert intuition, which, conversely, provides valuable design insights for the optimization community. We provide DesignX's Python project at~ https://github.com/MetaEvo/DesignX.

Figures (12)

• Figure 1: Left: Compared to manual design process, DesignX replaces human experts by two learnable agents. Right: Four dashed lines denote average performances of well-known human-crafted optimizers in decades. During pre-training, DesignX surprisingly discovers powerful optimizers superior to the ones crafted by human experts.
• Figure 2: Left: The dual-agent system in DesignX processes an optimizer workflow by the pre-order traversal of its program structure tree. Top Right: Agent-1 generates legal optimizer workflow in an auto-regressive fashion. Bottom Right: Agent-2 controls hyperparameters of the generated optimizer workflow by conditioning on the optimization progress information.
• Figure 3: The generalization performance of baselines on realistic scenarios.
• Figure 4: Left: Normalized importance factors of different module types for various problem characteristics. Right: Two look-into cases for interpreting design pattern learned by DesignX.
• Figure 5: Ratios of selected module types.