REVIEW: Learning the Preference of Bounded Agents / by Siobhán Cronin

O. Evans et al. Learning the Preference of Bounded Agents. NIPS Workshop 2015. Institute for the Future of Humanity.  

NOTE: While I love Squarespace for visual layout and platform compatibility, I am in the process of switching my blog to a platform that supports mathjax. Equation resplendence coming soon. 

This paper applies models of bounded and biased cognition to the creation of a generative model for human choices in decision problems. Perhaps more importantly, the authors attempt to infer preferences (not beliefs) from this model.

They focus their attention on four types of agents:

  • Hyperbolic-discounting (i.e. my favorite new way of saying procrastinating)
  • Agents using Monte Carlo approximations of expected utility
  • "Myopic" agents
  • Bounded value-of-information agents

Choice is made in proportion to a softmax function of expected utility:

Which relates to simulation of future internal states and choices, where U is how the agent computes immediate utility, M is how it predicts its future internal states, and h is how it modifies its state when it is passed to the choice function.