Kurzweil Prizes
These prizes, sponsored by KurzweilAI.net for the Fifth Conference on Artificial General Intelligence (AGI-12, Oxford, Emgland, December 8-11 2012), were awarded to recognize the Best Overall AGI Paper and the Best AGI Ideas of the conference. For the first time ever, the committee this year decided to award the Best AGI Idea prize to two papers.

The 2012 Kurzweil Prize for Best AGI Paper was awarded to Paul Rosenbloom for his paper entitled Deconstructing Reinforcement Learning in Sigma. Building upon his prior work on factor graphs, Sigma is the name the author now attributes to his graphical cognitive architecture. Among the author’s guiding principles in developing Sigma are a desire for “functional elegance” and an architecture based on fundamental guiding principles similar to Newton’s laws. In this article, the author demonstrates initial steps towards realizing the emergence of AGI behaviors from more basic primitives and their interactions. While more work still needs to be completed, this contribution demonstrates promise for Dr. Rosenbloom’s novel approach.

The 2012 Kurzweil Award for Best AGI Idea this year was split between two contributions: Helgi Helgason, Eric Nivel and Kristinn Thórisson for their article entitled On Attention Mechanisms for AGI Architectures: A Design Proposal; and Laurent Orseau and Mark Ring for their article entitled Space-Time Embedded Intelligence.

In their paper, Helgi Helgason, Eric Nivel, and Kristinn Thórisson tackle the important issue of attention management under the restriction of limited computational resources. Their proposed design includes both top-down goal-driven control as well as bottom up filtering; and is based on four basic requirements regarding the associated AGI architecture: that it be data-driven, use fine-grained processing units, have predictive capabilities, and use a unified sensory pipeline. The paper points out the inherent connection between the designs of attention management systems and the basic underlying AGI architectures, and provides important design characteristics for any AGI architecture.

The contribution by Laurent Orseau and Mark Ring extends the theoretical approach  toward defining intelligence taken by Legg and Hutter, by seeking a formal, sound, and practical theory. They consider agents that are fully embedded within their environment, subject to its space and time constraints,  and which are computed by and can be modified by this environment. Their framework allows agents to see their own source code and provides them the ability to predict source code changes. By considering a more realistic framework — one subject to an environment’s computational constraints — the formalization by Orseau and Ring addresses a long standing limitation of the approach of Legg and Hutter.