search

ksettaluri6 / AutoCkt

Deep Reinforcement Learning of Analog Circuit Designs
93 stars 36 forks source link

Questions about the choice of parameters initializations and normalization values #2

Open ahmedo42 opened 2 years ago

ahmedo42 commented 2 years ago

Hi, I wanted to understand some choices made in the code

from my understanding of the paper each parameter should be initialized to mid-range , but instead these "magic" ? values are used at every reset: https://github.com/ksettaluri6/AutoCkt/blob/2d01bc62cc6b26ec301d3598daac724c3b45af40/autockt/envs/ngspice_vanilla_opamp.py#L157

Also I understand that normalization is generally useful in RL , can you explain why these values are chosen for normalization? https://github.com/ksettaluri6/AutoCkt/blob/2d01bc62cc6b26ec301d3598daac724c3b45af40/eval_engines/ngspice/ngspice_inputs/yaml_files/two_stage_opamp.yaml#L15

can you explain the choice of these values and what effect this might have?

Thanks.