Active Feedback Learning with Rich Feedback

Hang Yu, Elaine Schaertl Short

Abstract: In this paper, we proposed rich feedback which contains multiple types of feedback to allow human teachers to provide a variety of useful information to the learning agent and modified Policy Shaping to accumulate the effects of rich feedback. Then we designed the ALPHA framework to actively request rich feedback and further developed it to use Deep Learning. The experimental results showed ALPHA with rich feedback can greatly improve learning and quickly lead the learning agent to the optimal solution.