A3C
is implemented as rlpack.actor_critic.a3c.A3C
. It inherits from A2C
class defined as rlpack.actor_critic.a2c.A2C
.
A3C implements the asynchronous Actor-Critic agent which supports gradient accumulation and mean reduction of gradients for training the model similar to A2C. Along with this A3C supports multiple actors which interact with the environment at the same time. Gradients of all actors' policy model are asynchronously reduced and mean gradient is populated for each actor's policy model before optimizer step. For more information refer here
Note that to run simulation with A3C, make sure to use rlpack.simulator_distributed.SimulatorDistributed, as this will run the simulation in the given environment in distributed setting.
Keyword: agent_name: "a3c"