Table of Contents

Table of Contents

Table of Contents

Table of Contents

Actor-Critic methods are a family of reinforcement learning algorithms that combine the benefits of value-based and policy-based approaches: an actor learns which actions to take, while a critic learns to judge how good those actions were and feeds that judgement back so the actor’s learning signal is far less noisy. This post builds the idea from the ground up: why pure policy gradients are noisy, how a baseline fixes that for free, and why the critic is really just a learned, state-dependent baseline.

Note: This post is created in conjunction with a discussion on chatbots performed as part of The Education Exchange, a series aimed at teachers to spread AI awareness in the hopes of encouraging further adoption. My hope is that this post highlights the gradual yet recently exponential explosion within the chatbot space and the ways in which these could impact educational institutions.