Rachel Lerch, RPI Graduate Student


Rachel Lerch, RPI Graduate Student

Carnegie 113

February 13, 2019 12:00 PM - 1:30 PM

Reinforcement Learning (RL) represents a family of algorithms concerned with goal-oriented learning based on interactions with the environment.  Recently, RL has been responsible for rapidly advancing the state of machine learning and AI systems (Silver et al. 2018). Despite these indisputable successes, some contemporary RL models still struggle to solve simple tasks which human players excel at (Mnih et al 2015). This raises the question whether these types of RL systems are plausible models for how animals including humans actually learn.


In this talk I will cover both seminal and recent developments in RL, conveying possible ways in which current insights from human cognitive science can be used in RL. Specifically, a particular emphasis of this talk will focus on how limits on information processing shape the kinds of learning algorithms we should be considering. I will describe early work that applies a new capacity-limited RL algorithm based on the mathematical framework of information theory. Compared to standard approaches, capacity-limited RL can achieve faster learning, and better generalization to novel environments. I will conclude by discussing future work in studying capacity limited RL in human systems. 



Add to calendar