An Inverse Risk-Sensitive Reinforcement Learning Approach to Learning Agent Preferences
Eric Mazumdar

Citation
Eric Mazumdar. "An Inverse Risk-Sensitive Reinforcement Learning Approach to Learning Agent Preferences". Talk or presentation, 24, August, 2017.

Abstract
The modeling and learning of human behaviors is becoming increasingly important as critical systems begin to rely more on automation and artificial intelligence. Yet, in this task we face a number of challenges, not least of which is the fact that humans are known to behave in ways that are not completely rational. We address the problem of learning a risk-sensitive agent's value function in a Markov decision process (MDP) through a gradient-based inverse risk-sensitive reinforcement learning approach. We model risk-sensitivity in a reinforcement learning framework by making use of models of human decision-making having their origins in behavioral psychology, behavioral economics, and neuroscience. We validate the approach on the classical Grid World MDP, and an MDP constructed from ride-sharing data.

Electronic downloads

Internal. This publication has been marked by the author for FORCES-only distribution, so electronic downloads are not available without logging in.

Citation formats

• HTML

  Eric Mazumdar. <a
  href="http://www.cps-forces.org/pubs/282.html"
  ><i>An Inverse Risk-Sensitive Reinforcement
  Learning Approach to Learning Agent
  Preferences</i></a>, Talk or presentation,  24,
  August, 2017.

• Plain text

  Eric Mazumdar. "An Inverse Risk-Sensitive Reinforcement
  Learning Approach to Learning Agent Preferences". Talk
  or presentation,  24, August, 2017.

• BibTeX

  @presentation{Mazumdar17_InverseRiskSensitiveReinforcementLearningApproachTo,
      author = {Eric Mazumdar},
      title = {An Inverse Risk-Sensitive Reinforcement Learning
                Approach to Learning Agent Preferences},
      day = {24},
      month = {August},
      year = {2017},
      abstract = {The modeling and learning of human behaviors is
                becoming increasingly important as critical
                systems begin to rely more on automation and
                artificial intelligence. Yet, in this task we face
                a number of challenges, not least of which is the
                fact that humans are known to behave in ways that
                are not completely rational. We address the
                problem of learning a risk-sensitive agent's value
                function in a Markov decision process (MDP)
                through a gradient-based inverse risk-sensitive
                reinforcement learning approach. We model
                risk-sensitivity in a reinforcement learning
                framework by making use of models of human
                decision-making having their origins in behavioral
                psychology, behavioral economics, and
                neuroscience. We validate the approach on the
                classical Grid World MDP, and an MDP constructed
                from ride-sharing data.},
      URL = {http://cps-forces.org/pubs/282.html}
  }
  

Posted by Carolyn Winter on 24 Aug 2017.
Groups: forces
For additional information, see the Publications FAQ or contact webmaster at cps-forces org.