Interactive Reinforcement Learning
Human Generated Reward
Presentation for Summer Camp 2015 May 25 2015
Reinforcement Learning
• Trial and error learning
• Explore and exploit
• Represent, predict and control
• Connect actions with rewards
• Maximize future reward
Sutton and Barto 1988
Interactive Machine Learning
Fails and Olsen Jr. 2003
Human Generated Reward
• Humans know more!
• Shaping systems to adapt
• Effectively reward learning
• Transfer learning through collaboration
• How can RL harness human reward?
Knox and Stone 2012
Kuhlmann et al. 2004
Learning from Advice Learning from Shaping
Blumberg et al. 2002
Thomaz et al. 2006
Learning from Demonstration
Left: Argall et al. 2010 Right: Koenemann et al. 2014
Learning from Trial and Error
Levine et al. 2015
Learning from Refinement
Cakmak et al. 2012
Application
• Shared control
• Augmented representation
• Integrate human and non-human interaction
• Autonomous prosthetics