RLgraph: Robust, incrementally testable reinforcement learning

We introduce RLgraph (GitHub repo), a RL framework decoupling logical component composition from deep learning backend and distributed execution. RLgraph manages session, variable/internal states, devices, scopes, placeholders, nesting, time and batchranks for each component and its incoming and outgoing dataflow. The core difference between using RLgraph and standard implementation workflows is that every component is fully specified explicitly: Its devices and scopes for computations and internal states (e.g. variable sharing) are explicitly assigned which spares developers the headaches of nested context managers.

Source: rlgraph.github.io