A Reinforcement Learning based Global Router
TimeWednesday, July 13th6pm - 7pm PDT
LocationLevel 2 Lobby
DescriptionMachine learning (ML) has been a common tool for solving many of the problems in the physical design of electronic circuits. However, most ML algorithms require a large amount of data to perform well. Reinforcement learning (RL), on the other hand, does not require a large amount of data to learn to perform a task. This research work discusses a framework to perform global routing with RL. Our proposed RL-based framework is built upon a collaboration-based approach where two players,
”Router” and ”Cleaner”, work together to improve an initially routed solution. The algorithm starts with an initial solution and an estimation of the locations of short routing violations. Then, Cleaner rips up the nets that are predicted to be causing the violations and the Router finds new paths for these nets. Hence, Router and Cleaner cooperate to generate an improved solution. It should be mentioned that no external human intervention or data is needed in this framework. For small circuits on a 5x5 routing grid, this approach successfully resolves about 94% of overﬂow-based short violation in 100 test netlists.