ZhouWenjun2019 / NARMA_DNN_RNN_RC

NARMA10 with DNN/RNN/RC

NARMA10 task

$$y(t+1)=0.3 y(t)+0.05 y(t)\left(\sum_{i=1}^9 y(t-i)\right)+1.5 u(t-9) u(t)+0.1$$

Results:

• RC: ~0.15

• DNN

  Input nodes	NMSE
  10	0.201
  20	0.023
  15	0.036
  11	0.093

• RNN

  Time steps	NMSE
  10	0.0731
  20	0.072
  15	0.0759
  11	0.0612

Something interesting:

• In RNN/DNN, it's better to set the output size 1;
• In RNN, it's better to keep the hidden states to the next sample in each phase;
• In DNN, it's better to set the input dim bigger than 10, otherwise, the performance will be worse;
• Usually, the performance of RC is worse than RNN/DNN;
• In my script, the performance of RNN is worse than DNN, and I didn't look into it.