When: February 14, 2014 from 11 AM to noon
Where: room W01
Speaker: Sergio Valcarcel
Title: Diffusion strategies for cooperative reinforcement learning
Abstract: We will introduce diffusion strategies and apply them to develop fully-distributed cooperative reinforcement learning solutions, in which agents in a network communicate only with their immediate neighbors to improve predictions about their environment. These algorithms are especially beneficial for off-policy sampling (meaning that the agents can learn to predict the response to a behavior different from the actual policies they are following): when the individual agents can estimate the solution, cooperation increases stability and reduces bias and variance of the prediction error; but, more importantly, the network is able to approach the optimal solution even when none of the individual agents could (e.g., when the individual behavior policies restrict each agent to sample a small portion of the state space). If time permits, we will sketch a mean-square-error performance analysis that establishes convergence under constant step-size updates.
