Sep 29

CS Tea - Paul Schrater Talk

Thu, September 29, 2016 • 4:00pm - 5:00pm (1h) • CMC 209

Learning and representing value in an uncertain world: Probabilistic models of value 

Paul Schrater, University of Minnesota 

While it is fair to say we choose what we value, the relative ease with which we make choices and actions masks deep uncertainties and paradoxes in our representation of value. For example, ambiguous and uncertain options are typically devalued when pitted against sure things - however, curiosity makes uncertainty valuable. In general, ecological decisions can involve goal uncertainty, uncertainty about the value of goals, and time/state-dependent values. When a soccer player moves the ball down the field, looking for an open teammate or a chance to score a goal, the value of action plans like passing, continuing or shooting depends on conditions like teammate quality, remaining metabolic energy, defender status and proximity to goal- all of which need to be integrated in real time. In this talk, we explicate two challenging aspects of human valuation using hierarchical probabilistic value representations. Hierarchical probabilistic value representations provide a principled framework for complex, contextual value learning and for the conversion of different kinds of value by representing more abstract goals across a hierarchy.  We show that preference reversals can be generated from rational value learning with hierarchical context, including anchoring and similarity effects, and we use our theory to experimentally induce preference reversals by manipulating subject’s history of experience. We also show how probabilistic representations of value can solve the problem of converting and integrating heterogeneous values, like metabolic costs vs. scoring a soccer goal. By modeling values in terms of probabilities of achieving better outcomes, we can integrate probabilistic value representations seamlessly into control theoretic models by decomposing complex multi-goal problems into weighted mixture of control policies, each of which produces a sequence of actions associated with more specific goal. Critically, the weights are inferences that integration all the time-varying probabilistic information about the relative quality of each policy. We use the approach to give a rational account for a set of reaching and oculomotor experiments with multiple goals.

Event Contact: Sue Jandro

Event Summary

CS Tea - Paul Schrater Talk
  • Intended For: Students, Faculty, Staff

+ Add to Google Calendar

Return to site Calendar
Go to Campus Calendar