Multi-Objective Optimization of Analog Circuits Using Reinforcement Learning

Abstract

Analog circuit design requires balancing conflicting performance metrics, making an accurate Pareto-Optimal Front (PoF) essential for trade-off analysis. Meta-heuristic algorithms are commonly employed to explore the PoF; however, the computational complexity involved in PoF exploration and the tendency to converge on suboptimal designs remain significant challenges that have yet to be fully addressed. This work introduces Multi-Objective Deep Deterministic Policy Gradient (MODDPG) for efficient PoF extraction in analog circuit design. By combining Reinforcement Learning (RL) with multi-objective optimization, MODDPG navigates high-dimensional design spaces more effectively than the conventional methods. It leverages a continuous action space and a tailored reward function to iteratively refine the design parameters. To validate the performance of the proposed approach, several MOO benchmark problems have been successfully solved. Then, experiments on analog circuit benchmarks have been perfromed to demonstrate that MODDPG outperforms traditional methods in key metrics, notably maximizing the Figure of Merit (FoM) and significantly advancing automation in circuit optimization. © 2025 Elsevier B.V., All rights reserved.