pascalwhoop / broker-python Goto Github PK

TensorForce, while extremely powerful and versatile, is a littl buggy. It's better than keras-rl because it offers more flexibilty (through observe_atomic) but is still showing issues.

YARL is supposed to change that. watch and reevaluate

tensorforce supports DQFD, which can be used to use another agent to teach us what to do. but, this needs either parasite learning (i.e. agent knows everything the other agent knows) or some other form of allowing them to have the same input which is hard because the input partially is the forecast which is internally generated.

Allow Agent to send PauseRequest to server and hook into debugging loop with python

• tests
• send message to server
• hook into debugger and send when debugger is hit
• send continue request on continue debugger

should it be learn/compete or maybe component based wholesale/demand/tariff ?

• Make Agent inherit from TensorForce agents
• extract preprocessing to separate
• create runner that works for offline learning