Automated Design of Complex Analog Circuits with Multiagent based Reinforcement Learning
DescriptionThe design search space explosion is a key challenge for complex analog circuits design automation. Inspired by multiagent planning theory, we propose three techniques to overcome this issue. Particularly, we (i) partition the circuit into sub-blocks based on circuit topology information and effectively reduce the complexity of design search space; (ii) introduce multiagent-based reinforcement learning and the interactions between agents mimic the design tradeoffs between circuit sub-blocks; (iii) utilize the twin-delayed technique to achieve training stability. Experiments on two analog circuit topologies and two technology nodes demonstrate that multiagent based RL algorithms can achieve better FoM and faster design speed.
Event Type
Research Manuscript
TimeTuesday, July 11th4:10pm - 4:25pm PDT
Location3002, 3rd Floor
Analog CAD, Simulation, Verification and Test