AI Learns to Match Optimizers to Complex Problems

In the world of artificial intelligence, solving large-scale optimization problems—those with over 1,000 variables—is a critical with applications ranging from designing electric aircraft to optimizing wind turbines. Traditionally, researchers have relied on a called Cooperative Coevolution (CC), which breaks down these massive problems into smaller, more manageable subproblems. However, a new class of problems, known as Heterogeneous Large-Scale Global Optimization (H-LSGO), has emerged, where subproblems vary not only in size but also in their mathematical landscapes. This heterogeneity, exemplified by complex real-world tasks like satellite design involving seven distinct modules, has rendered traditional CC approaches ineffective, leading to significant performance drops and increased runtime. A team from South China University of Technology has developed a novel solution: the Learning-Based Heterogeneous Cooperative Coevolution Framework (LH-CC), which uses reinforcement learning to dynamically select the most suitable optimizer for each subproblem, addressing this mismatch and paving the way for more efficient AI-driven optimization in diverse fields.

The core of this research is that dynamic optimizer selection is a pivotal strategy for solving complex H-LSGO problems. The researchers found that when subproblems exhibit diverse dimensions and distinct landscapes—such as mixing high-dimensional and low-dimensional components or different function types—using a single, fixed optimizer fails to navigate this heterogeneity effectively. In contrast, LH-CC formulates the optimization process as a Markov Decision Process, where a meta-agent learns to adaptively choose from a pool of optimizers, including both high-dimensional and low-dimensional options, for each subproblem. This approach allows the system to match the optimizer to the specific characteristics of the subproblem, such as its dimensionality or separability, leading to superior solution quality. For instance, on 3000-dimensional problems with complex coupling relationships, LH-CC achieved significantly better than state-of-the-art baselines, as detailed in the paper's experimental .

Ology behind LH-CC involves a sophisticated reinforcement learning setup designed to handle the intricacies of H-LSGO. The framework models the optimization process as a Markov Decision Process, with states that capture three categories of features: problem attributes (like dimensionality and separability), population dynamics (such as diversity and landscape ruggedness), and optimization progress (including computational budget usage and improvement ratios). These states, summarized in Table 1 of the paper, provide the meta-agent with comprehensive information to make informed decisions. The action space consists of a pool of candidate optimizers, and the agent uses an Actor-Critic network architecture, illustrated in Figure 2, to select the best optimizer at each step. To train the agent, the researchers employed Proximal Policy Optimization, with a reward function based on logarithmic improvements in solution quality, ensuring stable learning across diverse problem instances. Additionally, the team introduced a flexible benchmark suite, Auto-H-LSGO, which automates the generation of diverse H-LSGO problems, addressing the scarcity of test benchmarks in this domain.

From extensive experiments demonstrate LH-CC's effectiveness and efficiency. On a suite of 3000-dimensional H-LSGO problems generated using Auto-H-LSGO, LH-CC outperformed comparison algorithms across multiple metrics. As shown in Table 2, LH-CC achieved superior optimization performance, with significant advantages over baselines like OEDG-CMAES and OEDG-MMES-CMAES, which struggled with heterogeneity. For example, on problems like He2 and He3, LH-CC reached objective values orders of magnitude better than competitors, as visualized in Figure 4. In terms of efficiency, Table 3 reveals that LH-CC reduced runtime compared to traditional CC s, with runtimes around 1100 seconds for type 1 problems versus over 13000 seconds for some baselines, highlighting the penalty of optimizer-subproblem mismatch. Ablation studies, summarized in Figure 5, further confirmed that dynamic selection is crucial, as LH-CC outperformed fixed or random selection strategies, with normalized performance scores of 0.967 compared to 0.756 for random selection.

Of this research are substantial for real-world applications where optimization problems are inherently heterogeneous. By enabling AI to dynamically match optimizers to subproblems, LH-CC can improve efficiency in fields like satellite design, neuroevolution, and renewable energy systems, where complex, multi-module tasks are common. The framework's robust generalization, demonstrated across varying problem instances, optimization horizons, and optimizer families, suggests a 'train-once, apply-widely' paradigm that could reduce computational costs and expert intervention. For everyday readers, this means faster and more accurate solutions to large-scale problems, potentially accelerating innovations in technology and engineering. The Auto-H-LSGO benchmark also provides a valuable tool for researchers to test and develop new algorithms, fostering further advancements in the field.

Despite its successes, the study acknowledges several limitations. The framework currently relies on precomputed decomposition , which may not adapt to all problem structures dynamically. The training process, conducted with a reduced function evaluation budget of 1E+06 to mitigate computational costs, though shown to generalize to 3E+06, might not scale seamlessly to even larger budgets or more extreme heterogeneities. Additionally, the context memory mechanism for warm-starting optimizers, while effective, is a proof-of-concept that may require further refinement for broader applicability. The paper notes that future work could explore subspace contribution-based resource allocation or advanced network architectures to enhance autonomy. These limitations highlight areas for improvement but do not diminish the framework's pioneering role in addressing H-LSGO s.