This is going to be a meta branch to discuss any deep reinforcement learning implementations of algorithms to write, as well as any structural changes to the folder structure if people want to work on this and things break.
an unused reporting server in exe (ideally to kick off a haskell-variant of tensorboard)
the beginnings of some mypy type stubs for certain libraries (although mypy sometimes fails)
a3c (py)
ddpg (py)
dqn (py, in progress tensorflow-hs)
deep recurrent q-nets (py)
a model-based network (py)
a q-network (py, tensorflow-hs, backprop)
reinforce (py, in-progress tensorflow-hs)
In a different repository, there exists tensorflow-less versions of:
a2c
actor-critic
CEM
reinforce
So there are a few things to do here. One is to finish porting the python to haskell variants (tensorflow-haskell, and backprop -- optionally grenade, but I've had it blow up in my face already). The other is to implement more python algorithms, but these will be subject to certain constraints since the ultimate goal is to port to haskell. To hit parity with rllab*, we would still need:
This is going to be a meta branch to discuss any deep reinforcement learning implementations of algorithms to write, as well as any structural changes to the folder structure if people want to work on this and things break.
Currently, in deep-rl, there exists:
exe(ideally to kick off a haskell-variant of tensorboard)In a different repository, there exists tensorflow-less versions of:
So there are a few things to do here. One is to finish porting the python to haskell variants (tensorflow-haskell, and backprop -- optionally grenade, but I've had it blow up in my face already). The other is to implement more python algorithms, but these will be subject to certain constraints since the ultimate goal is to port to haskell. To hit parity with rllab*, we would still need:
* hitting parity with rllab would be nice since it would allow us to benchmark with results from Yan Duan, Xi Chen, Rein Houthooft, John Schulman, Pieter Abbeel. "Benchmarking Deep Reinforcement Learning for Continuous Control".