Online Learning RTL Synthesis for Automated Design Space Exploration

Black-box functions modeling is crucial in many optimization problems. However, the number of function evaluations is often limited by time or cost, as in Design Space Exploration (DSE). The literature mainly exploits genetic algorithms that, while effective in solving DSE, require several calls of expensive fitness functions (FTFNs), such as synthesis and place&route. Therefore, approximating these black-box calls would incredibly impact the final results. For these reasons, this manuscript proposes several improvements to existing DSE methodologies leveraging online learning models for an open-source approach generally applicable to FTFNs approximation. Through Adaptive Hoeffding Trees as surrogate models and Multi-Armed Bandits as controllers, our approach decides at each iteration whether to evaluate the real function or the surrogate model by annotating the FTFN. The proposed approach achieves a top speedup of 2.67x over the non-approximated DSE while delivering solutions comparable with the Pareto-optimal ones.