jimenius / internrl Goto Github PK

Implement DQN in Pytorch

Put configurations in a file is neat.

Random initialization for discrete state / action values to replace Exploring start.

DQN: Y = r + gamma * max(Q'(s'))
Double DQN: Y = r + gamma * Q'(s', argmax(Q(s')))

Since the program is configured by YAML, it is not necessary to have a scripts folder.

Network updates with number of steps rather than epochs

Synchronous update of PI and VI

Construct our own API to avoid using Gym API

https://docs.python.org/3.5/whatsnew/3.5.html#pep-484-type-hints

Show env key stats in logger

Current state and action dimension does not support Box type (in RLAgents/core.py)

Set parameters from external sources.
optimizer, learning_rate, verbose

eval is a Python keyword. Need to change it.

Configurations need logger.

Use Pytest to test programs

Use Average Q (V) instead of Average cumulative reward for learning evaluation

Construct networks from YAML file.

Decoding cfg in the main program doesn't seem to be efficient and tidy.

Currently main program has redundant part for different agents.

Implement DQN in TensorFlow Keras

So the asyncronous evaluation does not depend on number the position.