Symbolic Systems Forum - Dan Shapiro, Institute for the Study of Learning and Expertise

Dan Shapiro, Institute for the Study of Learning and Expertise and Center for the Study of Language and Information:

Value Driven Agents and the "Be all you can be" Guarantee

***note time change***

ABSTRACT:

Why don't we live in smart houses that automatically brew our coffee in

the morning, and play our favorite music at night? Why don't we drive

on automated freeways that get us there faster, even in rush hour? Why

won't NASA let the Mars rover avoid rocks on its own? Rather than blame

an absence of technology, I treat the issue as a lack of trust due to

poor communication. In this view, the goal is to give an agent a better

appreciation of its user's values, so that its choices won't have

unintended effects.

This talk will introduce a framework for constructing Value Driven

Agents that are guaranteed to maximally please their users within the

limitations of their given skills. I will begin by introducing the

concept of user-agent value alignment, and then outline a design

methodology and an implementation model for creating Value Driven

Agents of this kind.

BIO:

Dr. Daniel Shapiro is a Senior Researcher at Stanfordýs Center for the

Study of Language and Information and the Assistant Director of the

Institute for the Study of Learning and Expertise. Dr. Shapiroýs

current research interests focus on the design, development, and

application of autonomous agents that are motivated by a core concept

of value, as well as on the underlying technologies of decision theory,

reactive computing, and reinforcement learning.