What could we do if, instead of coding programs from the ground up, we could just specify the rules for a task, the success criteria, and make AI lear

Reinforcement Learning for Beginners - Maze Solver with Sarsa

submited by
Style Pass
2022-09-22 00:00:02

What could we do if, instead of coding programs from the ground up, we could just specify the rules for a task, the success criteria, and make AI learn to complete it?

Imagine the blessings humanity could unlock if we could automate all the tasks nobody wants to do, all the unsafe, unhealthy or uninspiring jobs. Or the ways Scientific progress could be sped up if big parts of the research process were accelerated.

That powerful question motivates Reinforcement Learning. Instead of programs that classify data or attempt to solve narrow tasks (like next-token prediction), Reinforcement Learning is concerned with creating agents, autonomous programs that run in an environment and execute tasks.

The algorithms that train a Reinforcement Learning agent are very hands-off compared to other branches of Machine Learning: just provide the agent with features describing the environment (like the graphics for a game, or other users’ recent posts for social media content), and give it rewards according to its actions.

This may not always work. At each time step (however we define our time steps), the agent will have to pick one of all possible actions, which can potentially form a huge universe of possibilities. There could also be delays between the time the action is chosen and when we learn if it was a good choice, or the environment could be hard or expensive to simulate –imagine simulating trades in the stock market, or Robotics tasks that deal with expensive equipment.

Leave a Comment