State-dependent valuation learning: fitness, state, reinforcement and choice.

Speaker:
Alex Kacelnik

Abstract:

Behavioural ecologists model choice as a function of fitness gains, while associative learning researchers model learning as a function of reinforcement. While the former typically by-pass the issue of how the subject learns which option confers higher fitness gains, the latter typically ignore the functional implications of learning mechanisms, or even how associative strength translates into choice. I try to bridge this gap with models and experiments in starlings and locusts that assume that preference at the time of choice is a function of state improvement at the time of learning the properties of each option. It is my guess that in AI an interesting and substantial question must be the choice of reinforcement criteria for any semi-autonomous decision-making robot, and the formulation of a built-in value function to occupy the role of fitness maximization in living beings.

References:

Pompilio, L., Kacelnik, A., & Behmer, S.T. (2006). State-dependent learned valuation drives choice in an invertebrate. Science 311: 1613-1615. Online: (pdf).