DesignX: Automated BBO Algorithm Design

• DesignX is an end-to-end automated framework that designs black-box optimization algorithms via meta-learning across synthetic and real tasks.
• It decomposes algorithm design into generating workflow structures from a modular library and dynamically tuning hyperparameters using a dual-agent RL approach.
• DesignX consistently outperforms traditional optimizers in benchmarks such as protein docking, AutoML, and UAV path planning, revealing novel algorithmic patterns.

DesignX is an end-to-end automated algorithm design framework for black-box optimization (BBO), simultaneously learning both the workflow structure and parameter schedules of optimization algorithms. By meta-learning across large distributions of synthetic and real tasks, DesignX autonomously generates novel algorithms that consistently outperform state-of-the-art human-crafted and meta-learned optimizers by orders of magnitude in diverse BBO settings, including synthetic benchmarks, protein docking, automated machine learning (AutoML), and UAV path planning (Guo et al., 23 May 2025).

1. Problem Formulation and Sub-Task Decomposition

DesignX targets the challenge wherein designing effective BBO algorithms is hindered by the lack of problem-specific insights and the labor-intensive, months-long process of manual algorithm construction. The algorithm design process is decomposed into two core sub-tasks:

1. Algorithm Workflow Structure Generation: Determining the ordered sequence of algorithmic modules (e.g., initialization, mutation, crossover, selection) that form a coherent optimizer.
2. Hyperparameter Control: Dynamically tuning numerical parameters for each module (e.g., mutation scale, crossover rate, inertia weight, adaptive schedules) as the optimization unfolds.

This decomposition motivates a dual-agent machine learning approach, allowing systematic exploration of a combinatorial space of algorithm structures and control regimes.

2. Modular Algorithmic Space: Modular-EC Library

Central to DesignX is the Modular-EC library, a comprehensive polymorphic space comprising 116 module variants spanning 10 types, distilled from decades of evolutionary computation research. Modules are categorized as uncontrollable (fixed function, no hyperparameters) or controllable (with tunable hyperparameters):

Module Category	Count	Key Examples
Initialization	5	Uniform, Sobol, LHS, Halton, Normal
Niching	3	Random, Ranking, Distance
Boundary_Control	5	Clip, Reflect, Periodic, Random, Halving
Selection	6	DE-like, Crowding, PSO-like, Tournament, etc.
Population_Reduction	2	Linear, Non-Linear
Restart_Strategy	4	Stagnation, Objective_Convergence, etc.
Mutation	49	DE variants, Gaussian, Polynomial, composites
Crossover	17	Binomial, Exponential, SBX, multi-strategy
Other_Update	10	PSO, FDR_PSO, CMA-ES, MMES, multi-strategy
Information_Sharing	1	–

Module variants encode both a unique 16-bit identifier and "topology_rule" constraints that restrict legal workflow sequences (e.g., mutation must precede crossover). Legal workflows are auto-regressively generated to respect these topology constraints.

3. Dual-Agent Reinforcement Learning Framework

DesignX employs a pair of cooperating RL agents, each implemented as Transformer (GPT-2) policies, to jointly optimize workflow structure and hyperparameter schedules:

• Agent-1 (Workflow Generator, $\pi_\phi$): Receives a 13-D problem feature vector $\mathbb{F}_p$ (4 basic + 9 Exploratory Landscape Analysis features), and autoregressively samples a valid sequence of module identifiers $\mathcal{A}_p$ to construct a workflow. The agent’s policy:

$$P(\mathcal{A}_p^m | …) = \text{Softmax}\left(\text{mask}(\mathcal{A}_p^{m-1}) W_{sample}^T H^{(m-1)}\right)$$

(where $H^{(m-1)}$ is the hidden state after $m-1$ GPT-2 blocks).

• Agent-2 (Hyperparameter Controller, $\pi_\theta$): At each optimization step $t$, Agent-2 receives a 9-D optimization state $\mathbb{O}_t$ encoding population statistics and budget. Conditioning on the workflow $\mathcal{A}_p$ and $\mathbb{F}_p$0, it outputs a Gaussian for each parameterized module's settings. Sampling:

$\mathbb{F}_p$1

The joint training objective treats both agents as parts of a joint MDP, with rewards based on normalized objective progress:

$\mathbb{F}_p$2

Agent-1 is updated via REINFORCE on final episode reward, Agent-2 is optimized with PPO on per-step rewards.

4. Large-Scale Meta-Training and Evaluation Protocol

DesignX is meta-trained on 9,600 synthetic BBO problems sampled from a base of 12,800 diverse problems:

• Base Functions: 32 canonical benchmark functions (e.g., Sphere, Rosenbrock, Rastrigin, Gallagher).
• Problem Modes: "Single" (one function), "composition" (weighted sum of 2–5 functions), and "hybrid" (variable segmentation among multiple functions).
• Randomization: Dimensionality $\mathbb{F}_p$3, search range $\mathbb{F}_p$4, budget $\mathbb{F}_p$5, along with shifts and rotations.

The split comprises 9,600 training and 3,200 test instances. Training proceeds for 100 epochs for both agents (Agent-1 batch size 128; Agent-2 PPO with 10-step rollouts, 3 epochs), executed on dual Xeon CPUs.

5. Quantitative and Qualitative Performance Analysis

Synthetic Benchmarks: On 3,200 held-out synthetic tasks, DesignX attains the lowest normalized average objective in 19/20 representative cases, with a mean value of $\mathbb{F}_p$6, outperforming the best baseline (MetaBBO and manually crafted) by over an order of magnitude ($\mathbb{F}_p$7).

Realistic Applications: In out-of-distribution evaluations:

• Protein docking (280 12-D instances): DesignX achieves native-like docking energies more than twice as quickly as CMA-ES and DE variants.
• AutoML (86 HPO-B tasks) and UAV path planning (56 30-D): DesignX surpasses all hand-designed and MetaBBO optimizers by significant margins.

Ablation and Scaling: Exclusive training of Agent-2 (hyperparameters only) yields collapse in performance, confirming the critical importance of structural design (Agent-1). Scaling studies suggest increased model capacity and larger task distributions further enhance capabilities, although even a single-layer GPT-2 on 10k tasks outperforms all baselines.

6. Algorithmic Insights and Design Patterns

Standardized module importance heatmaps reveal several trends:

• Mutation: Unimodal tasks favor exploitative single-strategy (e.g., DE/current-to-best/1), while multimodal tasks require composite mutation schemes.
• Population Reduction: More aggressive reduction is scheduled for tighter search ranges.
• Initialization: Demonstrates near-zero importance—any reasonable sampler suffices.
• Restart and Crossover: The discovered workflows often link novel restart logics with inventive crossover–mutation pairings.

DesignX automatically synthesizes previously unknown algorithm forms, including:

• Single-niche DE/current-to-rand/1+archive → Binomial (F=0.6, Cr=0.8) → Clip → DE-like selection, which excels on Rosenbrock-type problems.
• Multi-subpopulation Niching(Distance) with multi-mutation strategies and population reduction, effective on multimodal problems.

These schemata suggest the dynamic use of composite mutations for balancing exploration/exploitation and highlight the primacy of structure and population management over per-step parameter granularity.

7. Implications and Prospects

DesignX demonstrates that end-to-end learning of both workflow structure and parameter control produces optimizers surpassing decades of expert design, not only in aggregate metric performance but also by uncovering novel, effective patterns previously unexplored by the community. The dual-agent RL framework and the polymorphic Modular-EC space point towards a path for foundation models for algorithm design and motivate further research in automated scientific discovery and meta-learning for optimization (Guo et al., 23 May 2025).