Scalable Tools for Learning Models of Strategic Decision-Making
Lillian Ratliff

Citation
Lillian Ratliff. "Scalable Tools for Learning Models of Strategic Decision-Making". Talk or presentation, 28, May, 2015.

Abstract
The emergence of societal-scale cyber-physical systems such as the smart grid and intelligent transportation systems reveals new opportunities for efficiency yet exposes novel vulnerabilities. This efficiency-vulnerability tradeoff is a fundamental challenge facing S-CPS, wherein scarce resources must be allocated amongst competitive agents with misaligned goals. To manage this tradeoff, a coordinator can provide incentives to align these goals, for instance, by ensuring the equilibrium behavior optimizes a societal cost. In previous work, we derived an algorithm for synthesizing incentive strategies that lead to more efficient behavior when the preferences of the underlying agents are unknown to the coordinator and must be learned. In particular, we learned individual game-theoretic models of decision making. In this talk, we address the problem of learning models of strategic decision-making when the number of players is large. We leverage the individual-level game-theoretic model as a generative behavior model in a hierarchical Bayesian framework. Formulating higher-level estimation/classification tasks such as player segmentation or estimation of consumption of shared resources within in this framework, we provide bounds on inference error and optimal stopping time to reach a specified error tolerance. We show the results of computing these bounds using data from an experiment in which building occupants participated in a social game for inducing energy efficient usage of shared resources such as lighting.

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

  Lillian Ratliff. <a
  href="http://www.cps-forces.org/pubs/62.html"
  ><i>Scalable Tools for Learning Models of Strategic
  Decision-Making</i></a>, Talk or presentation, 
  28, May, 2015.

• Plain text

  Lillian Ratliff. "Scalable Tools for Learning Models of
  Strategic Decision-Making". Talk or presentation,  28,
  May, 2015.

• BibTeX

  @presentation{Ratliff15_ScalableToolsForLearningModelsOfStrategicDecisionMaking,
      author = {Lillian Ratliff},
      title = {Scalable Tools for Learning Models of Strategic
                Decision-Making},
      day = {28},
      month = {May},
      year = {2015},
      abstract = {The emergence of societal-scale cyber-physical
                systems such as the smart grid and intelligent
                transportation systems reveals new opportunities
                for efficiency yet exposes novel vulnerabilities.
                This efficiency-vulnerability tradeoff is a
                fundamental challenge facing S-CPS, wherein scarce
                resources must be allocated amongst competitive
                agents with misaligned goals. To manage this
                tradeoff, a coordinator can provide incentives to
                align these goals, for instance, by ensuring the
                equilibrium behavior optimizes a societal cost. In
                previous work, we derived an algorithm for
                synthesizing incentive strategies that lead to
                more efficient behavior when the preferences of
                the underlying agents are unknown to the
                coordinator and must be learned. In particular, we
                learned individual game-theoretic models of
                decision making. In this talk, we address the
                problem of learning models of strategic
                decision-making when the number of players is
                large. We leverage the individual-level
                game-theoretic model as a generative behavior
                model in a hierarchical Bayesian framework.
                Formulating higher-level estimation/classification
                tasks such as player segmentation or estimation of
                consumption of shared resources within in this
                framework, we provide bounds on inference error
                and optimal stopping time to reach a specified
                error tolerance. We show the results of computing
                these bounds using data from an experiment in
                which building occupants participated in a social
                game for inducing energy efficient usage of shared
                resources such as lighting.},
      URL = {http://cps-forces.org/pubs/62.html}
  }
  

Posted by Carolyn Winter on 10 Jun 2015.
Groups: forces
For additional information, see the Publications FAQ or contact webmaster at cps-forces org.