Learning Evolution via Optimization Knowledge Adaptation

TL;DR

OKAEM addresses the challenge of leveraging expanding knowledge bases in evolutionary algorithms by introducing a neural, attention-based framework that pre-trains on source tasks and adaptively tunes evolutionary operators during target-task optimization. Through learnable selection, crossover, and mutation, OKAEM achieves strong knowledge transfer, competitive self-tuning without priors, and efficiency gains across diverse benchmarks, including STOP, BBOB, and vision-language prompt tuning. The model demonstrates improved performance as prior knowledge accumulates and provides interpretability via learnable operator matrices, illustrating principles of natural selection and genetic recombination. This approach offers a scalable, transferable, and adaptable paradigm for complex optimization tasks in real-world settings.

Abstract

Evolutionary algorithms (EAs) maintain populations through evolutionary operators to discover diverse solutions for complex tasks while gathering valuable knowledge, such as historical population data and fitness evaluations. However, traditional EAs face challenges in dynamically adapting to expanding knowledge bases, hindering the efficient exploitation of accumulated information and limiting adaptability to new situations. To address these issues, we introduce an Optimization Knowledge Adaptation Evolutionary Model (OKAEM), which features dynamic parameter adjustment using accumulated knowledge to enhance its optimization capabilities. OKAEM employs attention mechanisms to model the interactions among individuals, fitness landscapes, and genetic components separately, thereby parameterizing the evolutionary operators of selection, crossover, and mutation. These powerful learnable operators enable OKAEM to benefit from pre-learned extensive prior knowledge and self-tune with real-time evolutionary insights. Experimental results demonstrate that OKAEM: 1) exploits prior knowledge for significant performance gains across various knowledge transfer settings; 2) achieves competitive performance through self-tuning alone, even without prior knowledge; 3) outperforms state-of-the-art black-box baselines in a vision-language model tuning case; 4) can improve its optimization capabilities with growing knowledge; 5) is capable of emulating principles of natural selection and genetic recombination.

Figures (6)

• Figure 1: Experimental results on 12 STOP benchmarks over 20 independent runs, each with a maximum of 5000 evaluations. a. Logarithmic plot of objective values (lower is better). b. Performance comparison between OKAEM, which learns different types of prior knowledge, and the corresponding source algorithms on target tasks. Data are presented as the mean and standard deviation of the objective values.
• Figure 2: Visualization of selection and mutation matrices learned by OKAEM on STOP1, STOP5, and STOP9. For the selection matrix, axes represent individuals ranked by their fitness scores, with 0 and 19 denoting the best and worst individuals, respectively. The matrix values are computed as the mean across all $H$ attention heads. For the mutation matrix, axes represent gene (decision variable) indices. a. STOP1: (i)-(ii) First and T-th generation selection matrices; (iii)-(iv) First and T-th generation mutation matrices. b. STOP5: (i)-(ii) First and T-th generation selection matrices; (iii)-(iv) First and T-th generation mutation matrices. c. STOP9: (i)-(ii) First and T-th generation selection matrices; (iii)-(iv) First and T-th generation mutation matrices.
• Figure 3: Experimental results on 24 BBOB Benchmarks (20 independent runs with random initialization, population size $N=20$, number of iterations $T=200$). a. Average convergence curves. b. Average GPU runtime. For more detailed performance comparisons, see Supplementary Fig. \ref{['figs1']}.
• Figure 4: Black-box prompt tuning for vision-language models. a. Illustration of the prompt tuning process. b. Average test accuracies on 8 common visual image classification tasks over five independent runs. c. Relative GPU runtime of the automatic tuning methods compared to OKAEM-ST.
• Figure 5: Sensitivity analysis of key parameters in OKAEM (20 independent runs with random initialization). The stopping criterion is set to a maximum of 5000 evaluations. a. Analysis of layers $L\in [1,2,4]$ and number of source tasks $K\in [10,50,100,1000]$. The heatmap displays the average normalized objective values across different types of STOP, with all values scaled to the $[0, 1]$ using min-max normalization. b. Population size $N\in[10, 20, 40, 50]$. The y-axis shows the normalized objective values. c. Dropout probability $p_C\in[0.3, 0.6, 0.9, 1]$ in crossover and dropout probability $p_M\in[0.3, 0.6, 0.9, 1]$ in mutation. The x-axis represents the objective values on a log scale.