Kary Främling: Reinforcement Learning with a Light-seeking Robot

In Reinforcement Learning (RL), an "agent" performs actions and observes 
received reward from the environment. This means that the agent actively 
has to explore the environment in order to collect training samples that 
it can learn from. Learning aims at identifying a "value function" that 
makes it possible to maximize the amount of reward received. 

For robotic applications, RL algorithms are usually tested in simulated 
environments. When using them in real-world applications, they often run 
into problems due to noise both in the environment and in the agent itself. 
In this presentation it is demonstrated that simple real-world robotic 
tasks can be treated using a linear model using artificial neural net (ANN) 
learning. The robot has to learn a sensorimotor mapping between three 
sensor inputs and five possible actions that allows it to reach a light source. 

The presentation includes some basic theory on RL and ANN, followed by 
a demonstration using a Lego Mindstorms robot.