SERS-3DPlace: Ensemble Reinforcement Learning for 3D Monolithic Placement

A novel Reinforcement Learning (RL) approach, that uses sequence-based placement representations and ensemble learning, is proposed for Monolithic 3D IC (M3D) placement. Our algorithm successfully chooses one of the best of the four types of placement perturbation actions most of the time. A New Ensemble-based policy allows to use multiple learning algorithms to choose good actions. RL produces an initial solution for Simulated Annealing (SA) that generates the final answer. To illustrate the effectiveness of SERS-3DPlace, we tested it on 8-128-bit MUX-based right arithmetic shifter (Muxs) circuits and a circuit with non-regular connections, as compared to Mux-based shifters, all implemented in 2-layer Monolithic 3D technology. The experimental results show that the Ensemble-based policy performs 2.5X better than the Multilayer Perceptron (MLP)-based policy, and the new SERS-3DPlace shows 2X improvement in the RL stage over RS3DPLace [1].