DQNCloud

A cloud deployment for DQN optimization network. Takes a POST to the /optimize endpoint containing previous/next state data, environment rewards, and trial/episode/step information.

/optimize

POST header should be application/json

POST body should be:

{
  "image": <list>,
  "reward": <list>,
  "step": <int>,
  "episode": <int>,
  "action": <list>,
  "last_state": <list>,
  "next_state": <list>,
  "inventory": <bool>,
  "version": <int>
}

/version

GET the current version of the policy DQN.

/policy?version=

GET this url to download a copy of the latest policy weights optimized by the network. It takes one query string parameter, version which specifies which policy net weights version file to download.

config.py

A config.py needs to be added to the directory. It should have the following constants defined:

CONVOLUTION_WIDTH = <int> # How wide will the initial convolution layer be?
CONVOLUTION_HEIGHT = <int> # How high will the initial convolution layer be?
BATCH_SIZE = <int> #how large of a batch will be processed by the DQN on optimization
GAMMA = <float> #The gamma to use in discounting rewards over time
DEVICE = <string> # pyTorch device (cpu or cuda)
TARGET_UPDATE = <int> #How many episodes before updating target DQN network
STATE_DICT_PATH = <string> #Server directory path for storing the state weight dictionary
REPLAY_MEMORY = <int> #Size of the replay memory
MONGO_CLIENT = <string> #ip:port of mongoDB instance
ACTION_SPACES = <list> #Array of integers containing the number of actions for each discreet action space
HOST = <string> #The IP of the host running this server