The design space for mixed-signal circuits topologies involves numerous connections that are difficult to comprehend. Besides, evaluating each topology incurs substantial computational costs. To address these challenges, analog/mixed-signal topology synthesis method based on reinforcement learning algorithm is proposed. Delta-Sigma analog-to-digital converter is selected as a design case. The proposed framework is composed of two level of optimization in a hierarchical way. The top-level algorithm decides the topology. At lower level of the framework, the reinforcement learning algorithm with graph attentional network is implemented to optimize each topology thus the design with lowest power consumption is found. The transfer learning is also implemented to improve the sample efficiency of optimization algorithm, where pre-trained RL agents can be deployed on similar other topologies.
24/5/2024 11:00 - 12:00
ESAT Aula L
