Reinforcement Learning with TensorFlow and TF-Agents
TF-Agents is a library for reinforcement learning (RL) in TensorFlow (TF). It makes the design and implementation of various algorithms easier by providing a number of modular components corresponding to the core parts of an RL problem:
- An agent operates in an environment and learns by processing signals received every time it chooses an action. In TF-Agents, an environment is typically implemented in Python and wrapped in a TF wrapper to enable efficient parallelization.
- A policy maps an observation from the environment into a distribution over actions.
- A driver executes a policy in an environment for a specified number of steps (also called episodes).
- A replay buffer is used to store experience (agent trajectories in action space, along with associated rewards) of executing a policy in an environment; the buffer content is queried for a subset of trajectories during training.
The basic idea...