News

Paul Rosenbloom Wins Kurzweil Prize for Best Paper at Artificial General Intelligence Conference 2012

December 11, 2012

Paul Rosenbloom has once again been honored with a Kurzweil Prize at the annual Artificial General Intelligence Conference (AGI). This year, Rosenbloom, a computer science professor at the USC Viterbi School and a project leader at ICT, won the best paper award for his paper “Deconstructing Reinforcement Learning in Sigma”. This work is an extension of Rosenbloom’s paper at the 2011 AGI conference, which received the Kurzweil Prize for Best Artificial General Intelligence Idea.

Both prizes, the first for an early stage idea and the more recent one for the maturation of that work, recognize Rosenbloom’s pioneering efforts toward building computer systems that can behave like people in how they make decisions and solve problems.

Artificial general intelligence, or AGI, refers to the design of systems that can emulate full-range human intelligence as opposed to artificial intelligence systems that focus on modeling narrow or specific functions like generating speech, acquiring language or planning actions.

“Significant progress has been made in many individual areas since the founding of AI, but such progress by itself doesn’t yield human level intelligence,” Rosenbloom said. “AGI represents a return to this original vision of AI.”

Rosenbloom leads such an effort at ICT, where he is building the next generation of virtual human architecture – sort of a brain for computer-driven characters – that should enable them to behave appropriately when combined with the proper knowledge and skills. ICT is a leader in research and development of virtual humans, and it is hoped that Rosenbloom’s new architecture – when combined with other developments at ICT and elsewhere – will lead to more human-like systems.

Potential applications could include more intelligent virtual humans, robots and agents – virtual negotiation partners, for example, that learn from their mistakes, react to current situations and alter their behaviors depending on past and present interactions.

“The goal is to integrate all the mechanisms for thought, language, speech and motor control into a single system that can learn from experience,” he said. “Usually when you add functionality in architectures, you add complexity. My work is trying to simplify the process with an architecture that combines elegance with generality.”