Adaptive AE toy task

E2C is too complicated to deal with, so I simplified the model and am just using a simple autoencoder to do prediction and reconstruction. No dynamics or anything fancy.

Modified the potential function, with datasets deliberately occupying fewer samples.

Random policy

Adaptive policy

Note: the adaptive policy uses random noise, but it seems to perform poorly when sampleNormal $e \sim N(0,1)$ noise. Further investigation needed.


Couple notes to self:

• When there are issues with "need to feed placeholders", it's likely b.c. you forgot to set var_list in optimizer.
• These simple models train faster on CPU than GPU.