James, A., Kong, Z. H., Ata, S. K., Aung, K. M. M., Yeo, K. S., Foo, C. S., & James, A. (2025). The Initialization Effect: Refining Bayesian Optimization for Circuit Design. IEEE Transactions on Circuits and Systems for Artificial Intelligence, 1–10. https://doi.org/10.1109/tcasai.2025.3609742
Abstract:
Bayesian optimization (BO) has emerged as a powerful sample efficient technique for optimizing expensive and time-consuming design of analog circuits. The BO framework leverages probabilistic models to efficiently explore design spaces, allowing for a streamlined approach to identify optimal designs that meet specific performance targets. Traditionally, BO framework is initialized with predefined or randomly generated samples, thereby failing to fully exploit the latent insights concealed within the data leading to suboptimal results and lengthy convergence time. This paper delves into the significance of initialization in BO for analog circuit design, investigating its impact on the efficiency and effectiveness of the optimization process. The study demonstrates that the model can achieve optimal design outcomes using as few as 16 data points, highlighting the potential for data-efficient learning in complex design tasks. Through empirical analysis and experimentation, we explore various initialization strategies and their influence on optimization performance, considering factors such as convergence speed, solution quality etc. The results highlight the influence of initialization strategies, with approaches such as Latin hypercube sampling (LHS) and K-center greedy achieving faster convergence compared to other methods like K-means and random initialization. This study demonstrates how effective initialization strategies in BO can reduce the need for extensive data collection, by optimizing design outcomes with fewer iterations and samples. Furthermore, we elucidate the underlying mechanisms through which different initialization approaches affect the optimization process, providing insights into the interplay between initialization, exploration, and exploitation in analog circuit design optimization. Our findings shed light on the importance of thoughtful initialization strategies in harnessing the full potential of BO for analog circuit design, offering valuable guidelines for circuit designers.
License type:
Publisher Copyright
Funding Info:
This research / project is supported by the Agency for Science, Technology and Research (A*STAR) - Advanced Manufacturing and Engineering (AME) Programmatic Fund
Grant Reference no. : A20H6b0151
This research / project is supported by the Agency for Science, Technology and Research (A*STAR) - Advanced Manufacturing and Engineering (AME) Programmatic Fund
Grant Reference no. : A19E3b0099
This research / project is supported by the National Research Foundation - Competitive Research Programme
Grant Reference no. : NRF-CRP20-2017-0006