Artificial Intelligence Now

Chapter 6. Reinforcement Learning Explained

Junling Hu

A robot takes a big step forward, then falls. The next time, it takes a smaller step and is able to hold its balance. The robot tries variations like this many times; eventually, it learns the right size of steps to take and walks steadily. It has succeeded.

What we see here is called reinforcement learning. It directly connects a robot’s action with an outcome, without the robot having to learn a complex relationship between its action and results. The robot learns how to walk based on reward (staying on balance) and punishment (falling). This feedback is considered “reinforcement” for doing or not doing an action.

Another example of reinforcement learning can be found when playing the game Go. If the computer player puts down its white piece at a location, then gets surrounded by the black pieces and loses that space, it is punished for taking such a move. After being beaten a few times, the computer player will avoid putting the white piece in that location when black pieces are around.

Reinforcement learning, in a simplistic definition, is learning best actions based on reward or punishment.

There are three basic concepts in reinforcement learning: state, action, and reward. The state describes the current situation. For a robot that is learning to walk, the state is the position of its two legs. For a Go program, the state is the positions of all the pieces on the board.

Action is what an agent can do in each state. Given ...