search

vlad17 / mve

MVE: model-based value estimation
Apache License 2.0
10 stars 0 forks source link

[explanation] Set up DNN logging #17

Closed vlad17 closed 6 years ago

vlad17 commented 6 years ago

For dynamics and controller neural networks, set up (optional, flag-controlled) logging through logz.py for neural network average gradient magnitude (for every on-policy aggregation iteration). In addition, record the network's corresponding objective loss before / after training that iteration.

vlad17 commented 6 years ago

This is especially important to verify that the learner is converging appropriately. In addition to the raw iteration-to-iteration losses for the learner, it would be important to report the learner's loss on the VALIDATION dataset; that is, how the learner actually predicted the MPC's behavior.

vlad17 commented 6 years ago

(solved by #66 )