rshen36/dl-project

Project repo for Advanced Topics in Deep Learning S17

dl-project's Issues

ES on A3C gradients?

One (or few) steps A3C to compute gradients, perform several ES steps on gradients (instead of on parameters)?

Fix weird batch sizes issue

Figure out redis sockets and communication in general

Debug and test worker code

Redirect logging output to file

Fix weird episode length issue

Do full code refactoring/re-organization

Particularly, add bash/python script to encapsulate starting all processes into one command

Add visualizations for training

Implement next policy class

Implement a Policy subclass for a particular autonomous driving environment, most likely OpenAI Universe's DuskDrive environment

Debug/optimize hyperparameters

Implement convolutional, batch norm, pooling, etc. layers

Figure out how to run python modules in background

Get simple worker agent running in Gym Go

Port universe-starter-agent code to Gym

Combine parallelized ES with A3C

Take advantage of similar asynchronous implementations, just use some sort of hyperparameter to randomly decide which algorithm to use at each task?

Port code to Discovery

Port LSTMPolicy code to es_distributed

Implement lstm/gru layer?

Might help?

Simultaneous A3C and ES?

Separate workers doing A3C and ES updates on the same task?

Improve training visualizations

Would like to use TensorBoard instead of matplotlib
Would also like to do visualizations only after training for computational efficiency