The Fifth Elephant 2023 Monsoon

Reinforcement Learning (RL) is a subfield of machine learning that involves interacting with the environment to improve performance. Scientists have been using various games as a way to test and compare different learning and planning methods. Back in 1992, Gerry Tessauro used Reinforcement Learning to train a neural network to play Backgammon. Since then, similar techniques have been used to create agents that can play games like Chess, Go, Shogi, Diplomacy, Doom, Dota2 and Starcraft even better than humans can.

We’ll take a look at how these Reinforcement Learning algorithms have developed over time and how they’ve helped to create complex agents that can play games. I’ll share my experience of creating agents to play a collaborative cooking game called Overcooked at Unity, and the different challenges that come with training using Reinforcement Learning.

After we’ve explored how Reinforcement Learning has been used in games, we’ll discuss how it’s being used in real-world applications like Google Translate and chatGPT. We’ll also look at what the future might hold for improving the reasoning abilities of these large language models.

Key Points:

1. What is Reinforcement Learning?
2. A quick look at how Reinforcement Learning methods have evolved over time
3. My experience of training non-playing characters for the collaborative cooking game - Overcooked
4. How Reinforcement Learning is being used in applications like Google Translate and chatGPT
5. Future directions to improve the reasoning abilities of large language models