Abstract
To address the design problem where multiple indicators of the LC voltage-controlled oscillator (LCVCO) are mutually restrictive, as well as the issue of insufficient efficiency in traditional manual parameter tuning, an automatic parameter optimization method based on improved Q-learning is proposed. A circuit model was established based on the 55 nm CMOS process, converting parameter optimization into a decision-making process driven by states, actions, and rewards. Combined with local actions, boundary constraints, effective action filtering, exponential decay exploration rate, and restart mechanism in case of stagnation, the search efficiency under a limited evaluation budget is improved. Within 300 evaluations, the improved Q-learning achieved a comprehensive quality factor of -203.20 dBc/Hz, which is higher than the -197.91 dBc/Hz of the standard Q-learning. After optimization, the phase noise of the experimental circuit at a 1 MHz frequency offset improved from -96.85 dBc/Hz to -108.47 dBc/Hz, and the power consumption decreased from 2.99 mW to 2.61 mW. In 3000 evaluations, the improved Q-learning reached the highest reward on the 194th evaluation, earlier than simulated annealing and random search, and the number of evaluations reaching the optimal value was distributed between 170 and 235 times under the three random seeds. The results show that this method can improve the parameter search efficiency of LCVCO under a limited evaluation budget and has good optimization performance and repeatability.
Key words
LCVCO /
reinforcement learning /
Q-learning /
parameter optimization /
FoMT
Cite this article
Download Citations
An automatic optimization method for LCVCO circuit parameters based on improved Q-learning[J]. Integrated Circuits and Embedded Systems. 0 https://doi.org/10.20193/j.ices2097-4191.2026.0129
{{custom_sec.title}}
{{custom_sec.title}}
{{custom_sec.content}}