Evolutionary Generative Optimization: Towards Fully Data-Driven Evolutionary Optimization via Generative Learning

TL;DR

EvoGO addresses a core limitation of traditional evolutionary optimization by delivering a fully data-driven framework that learns the entire search process end-to-end via a paired generative architecture. It replaces handcrafted crossover/mutation with forward and inverse generative models trained against a landscape surrogate, guided by a composite loss that balances reconstruction, distributional alignment, and optimization direction. The approach achieves rapid convergence (often within 10 generations) and strong performance across numerical, control, and robotic benchmarks, with substantial GPU-accelerated speedups and favorable data efficiency. Ablation studies confirm the necessity of the surrogate, paired architecture, and carefully designed losses, while supplementary results highlight robustness and limitations, including data requirements and potential gains from offline training and transfer learning.

Abstract

Recent advances in data-driven evolutionary algorithms (EAs) have demonstrated the potential of leveraging historical data to improve optimization accuracy and adaptability. Despite these advancements, existing methods remain reliant on handcrafted process-level operators. In contrast, Evolutionary Generative Optimization (EvoGO) is a fully data-driven framework designed from the objective level, enabling autonomous learning of the entire search process. EvoGO streamlines the evolutionary optimization process into three stages: data preparation, model training, and population generation. The data preparation stage constructs a pairwise dataset to enrich training diversity without incurring additional evaluation costs. During model training, a tailored generative model learns to transform inferior solutions into superior ones. In the population generation stage, EvoGO replaces traditional reproduction operators with a scalable and parallelizable generative mechanism. Extensive experiments on numerical benchmarks, classical control problems, and high-dimensional robotic tasks demonstrate that EvoGO consistently converges within merely 10 generations and substantially outperforms a wide spectrum of optimization approaches, including traditional EAs, Bayesian optimization, and reinforcement learning based methods. Code is available at: https://github.com/EMI-Group/evogo

Figures (15)

• Figure 1: The overall framework of EvoGO. The historical dataset $\mathcal{D}$ is composed of $\{\mathcal{D}_0 \cup \cdots \cup \mathcal{D}_t\}$, where each $\mathcal{D}_t = \left\{ (\bm{x}_{t,i}, y_{t,i}) \mid 1 \le i \le N \right\}$ , with $\bm{x}_{t,i}$ and $y_{t,i}$ denoting the population and its fitness, respectively. In data preparation phase, the historical population is first selected to form the training set of current generation and augmented by an optional learning-based data augmentation strategy. Then, it is separated into the superior and inferior groups with smaller and larger fitnesses, respectively, before a Cartesian product is applied to generate the paired training dataset. In model training phase, the framework first trains a surrogate model solely for landscape estimation. Then, the surrogate model is frozen and combined with a pair of generative models to form a composite model for effective generative model training. Later, the composite model is trained via the proposed training loss tailored for comoplex optimization. In population generation phase, the inferior-to-superior generative model is used to generate the new solutions in parallel.
• Figure 2: Overview of the model architecture and composite loss functions $\mathcal{L}$ in EvoGO. The top, middle, and bottom rows illustrate the computational workflows for the similarity loss $\mathcal{L}_{\text{sim}}$, the reconstruction loss $\mathcal{L}_{\text{rec}}$, and the optimization loss $\mathcal{L}_{\text{opt}}$, respectively. The input pair $(\bm{p}, \bm{q})$ corresponds to solutions with inferior and superior fitness, respectively.
• Figure 3: The convergence curves of EvoGO and the compared baselines in standard settings. The $y$-axes for numerical functions are in logarithmic scale, while those for other tasks are in linear scale since their true optimal is unknown. One-sigma error bars are adopted since two-sigma ones result in excessively large error bars for some algorithms. Please refer to Supplementary Document E for the complete figure.
• Figure 4: The rewards of TPE, CMA-ES, PPO, and EvoGO on the Hopper environment provided by Brax. To highlight the advantage in runtime of large population sizes, the rewards are plotted versus the wall-clock time, while the FEs are indicated by different markers. We limited the number of FEs and runtime to 10,000 and one hour respectively for all algorithms.
• Figure 5: The convergence curves of EvoGO and its upper-bound variant "Real Eval", which directly utilizes real numerical functions instead of surrogate models.