How to solve Reinforcement Learning when there are ZERO rewards (Curiosity & RND)

In this video, we will learn about two great RL methods for self supervised exploration - Curiosity and Random Network Distillation (RND). We will use a popular on-policy RL algorithm (A2C or Advantage Actor-Critic) to do our exploration of this fascinating space. Along the way, we will learn some code examples in Python and Pytorch, and study why exactly these methods work and the various challenges they solve. Curiosity teaches agents to explore states that they find the least predictable, and RND teaches agents to explore states that are the most "novel". The neural network architectures behind the two are incredibly simple yet genius, and they avoid noisy reward hacking solutions for a more streamlined intrinsic rewarding approach.

#reinforcementlearning #ai #deeplearning

Follow on Twitter: https://x.com/neural_avb

Buy me a coffee at https://ko-fi.com/neuralavb !

To join our Patreon, visit: https://www.patreon.com/NeuralBreakdownwithAVB

Members get access to EVERYTHING behind-the-scenes that go into producing my videos. Plus, it supports the channel in a big way and helps to pay my bills.

Papers:

Curiosity: https://arxiv.org/abs/1705.05363

RND: https://arxiv.org/abs/1810.12894

A2C: https://arxiv.org/abs/1602.01783