Deep Learning Essentials

Actor-critic-based algorithms

Both value-based and policy-based algorithms act independently without learning from each other. Actor-critic-based algorithms aim to improve upon this drawback. This algorithm combines value functions with policy iteration algorithms, so that the policy is updated at every iteration based on the updated value function. Figure, Actor-critic-based reinforcement learning, illustrates this workflow in more detail:

Actor-critic-based reinforcement learning

In the preceding figure, you start with an initial policy by an Actor (policy-based algorithm). The Critic or the value function receives the new state as well ...