Reinforcement Learning: An Introduction

http://webdocs.cs.ualberta.ca/~sutton/book/ebook/the-book.html

by fogus • 16 years ago

20 4 16 years ago

webdocs.cs.ualberta.ca

bioweek• 16 years ago

Could I train a robotic hand to catch a ball with this? What would be involved?

mahmud• 16 years ago

That might require you to throw the ball in every possible combination, and capture a high resolution time series of its positions, in order to train the AI. Taking into account all the various parameters.

There will also be some heavy image processing, in your motion detection engine.

I recommend you start with a computer screen, and train your robot to follow the motion of a virtual ball on the screen. That way you can control every part of the experiment without a heavy investment in equipment (or all the pesky details of the real analog world.)

the_real_r2d2• 16 years ago

In theory yes you could. RL is based on trial and error. The agent senses the environment in the form of states (ligh=state1, dark=state2) and it performs actions according to a policy to receive positive or negative rewards. For example your goal is to have a room always lighten, the agent will sense the environment (it is dark) it will turn the lights on and it will receive a positive reward. Eventually the agent will learn that in order to receive positive rewards it needs to turn on the light when it is dark and turn it off when it is light. Actually RL it is more complex than that but that is the basic idea. If you are interested go to scholar.google.com and search for papers about reinforcement learning. Important authors are Barto, Sutton, Watkins (Q-Learning), Littman, Kudenko, Stone, Clauss, etc.

chromophore• 16 years ago

This has been here before : http://news.ycombinator.com/item?id=795388

Reinforcement Learning: An Introduction

Related Stories