ray/rllib/agents/a3c

Advantage Actor-Critic (A2C, A3C)

Overview

Advantage Actor-Critic proposes two distributed model-free on-policy RL algorithms, A3C and A2C. These algorithms are distributed versions of the vanilla Policy Gradient (PG) algorithm with different distributed execution patterns. The paper suggests accelerating training via scaling data collection, i.e. introducing worker nodes, which carry copies of the central node's policy network and collect data from the environment in parallel. This data is used on each worker to compute gradients. The central node applies each of these gradients and then sends updated weights back to the workers.

In A2C, the worker nodes synchronously collect data. The collected data forms a giant batch of data, from which the central node (the central policy) computes gradient updates. On the other hand, in A3C, the worker nodes generate data asynchronously, compute gradients from the data, and send computed gradients to the central node. Note that the workers in A3C may be slightly out-of-sync with the central node due to asynchrony, which may induce biases in learning.

Documentation & Implementation:

1. A2C.

   Detailed Documentation

   Implementation

2. A3C.

   Detailed Documentation

   Implementation