Micro-architecture design space exploration (DSE) is essential for the optimization of processor performance. However, existing approaches generally fall short in interpretability, preventing designers from comprehending and fine-tuning decisions. In this article, we develop an Explainable Fuzzy Neural Network (FNN) framework coupled with multi-fidelity Reinforcement Learning (RL) for micro-architecture optimization. Instead, this proposed approach is a fuzzy-based one. It induces readable design rules by mimicking a neural network without limiting its flexibility. We employ a multi-fidelity RL strategy to simultaneously enable fast evaluation on analytical models, as well as perform accurate design validation on register transfer level (RTL) simulations to save time and guarantee correctness. Opportune learning of this kind greatly lowers computation costs and retains solutions of suitable quality. The FNN learns interpretable decision rules in an unsupervised fashion, allowing the designer to visualize optimization paths taken and incorporate domain knowledge to guide exploration. Our experiments show that our approach substantially improves the efficiency and accuracy over state-of-the-art DSE methods, yielding near-optimal micro-architectures with a very small sample budget. The interpretability of the framework also enables designers to visually inspect and optimize architectural trade-offs. Nearly closes the gap between black-box optimization and human-guided decision-making, our methodology establishes a solid groundwork for both explainable and efficient micro-architecture DSE, offering a pathway for future work that turns to more transparent and modelled methods of processor design.
Alkhiri et al. (2026) studied this question.