Weight Agnostic Neural Networks

Luo Yi
Dec. 12, 2019

Pros & Cons of WANNs

Experiments with Reinforcement Learning Tasks

swingup

  1. Swing Up
  1. Biped
  1. CarRacing

Performance Along with Fixed Topology

swingup

>Weights\score\Swing Up:WANN,Fixed Topology
Random,57,21
Random Shared,515,7
Tuned Shared,723,8
Tuned,932,918
>Weights\score\Biped:WANN,Fixed Topology
Random,-46,-129
Random Shared,51,-107
Tuned Shared,261,-35
Tuned,332,347
>Weights\score\CarRacing:WANN,Fixed Topology
Random,-69,-82
Random Shared,375,-85
Tuned Shared,608,-37
Tuned,893,906

Why?

WANNs' weights are trained to be so.

  • Without training, it can do.
  • With little training, it's near optimal.

Parameters and HyperParameters

params-diff

But it is unfair.

WANNs' architectures are shrunk to be so.

Can WANN beat NAS? I wonder.

Page Not Found

Try to search through the entire repo.