It's actually not so gradiose.
Julia implementations of deep reinforcement learning algorithms. Uses the POMDPs.jl framework for representing (Partially Observable) Markov Decision Proccesses (which itself is a framework for describing sequential decision making problems).
Working-ish:
- Deep Q Learning (DQN)
Currently working on:
- Model-Free Episodic Control (QEC)
- Deterministic Policy Gradient (DDPG)
To work on:
- Stochastic Policy Gradient (SDPG)
- Trust Region Policy Optimization (TRPO)
- Trust Region Generalized Advantage Estimate (TRGAE)
- VIME
- Misc NN models for MXNet (GAN, VAE, RNN)
NOTE: the signature for solve doesn't exactly match POMDPs.solve. It is `solve(::Solver, ::MDP, ::Policy, ::rng)
NOTE: try to define your vectors as Float32 if possible (which is what mxnet uses)
Documentation someday