This reading group focuses on methods and algorithms that allow computational agents to learn from human teachers via natural modes of communication ("natural" will be more of a goal than a full constraint). Specifically, this can include learning from advice, demonstration, or reinforcement.

Potential benefits of learning from a human teacher include the following.

1. 1)These natural methods often require less training samples to reach good performance than than autonomous learning algorithms that require an evaluation function (i.e., a reward function or fitness function).
   

2. 2)A teacher without programming skills can transfer knowledge of a task to the agent or even define the goals of a task, allowing real-world agents to learn from lay users.
   

3. 3)Agents can learn in the absence of a coded evaluation function. Defining evaluation functions can be very difficult, especially when multiple goals must be balanced.
   

Meetings are typically held in the conference room in ENS 32NE.

Questions? Contact Brad Knox at bradknox*AT*cs.utexas.edu