Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Universal Intelligence: A Definition of Machine Intelligence

Shane Legg, Marcus Hutter

TL;DR

The paper tackles the fundamental challenge of defining intelligence in a way that applies beyond humans by formalizing a broad, environment-driven notion of machine intelligence. It derives a universal intelligence measure, Υ(π), grounded in a reinforcement-learning framework and weighted across all computable, reward-summable environments using an Occam-inspired prior 2^{-K(μ)} over environment complexity. It then analyzes the properties of this measure, discusses its connection to the AIXI agent, and surveys existing informal and formal definitions/tests of machine intelligence, highlighting strengths, limits, and practicable approximations. The work argues for a theory-grounded, broadly applicable, and potentially testable concept of machine intelligence, while acknowledging computability constraints and the need for practical evaluation protocols. Overall, it provides a rigorous foundation for understanding and comparing intelligent systems across diverse environments and motivates future work on implementable tests that approximate universal intelligence.

Abstract

A fundamental problem in artificial intelligence is that nobody really knows what intelligence is. The problem is especially acute when we need to consider artificial systems which are significantly different to humans. In this paper we approach this problem in the following way: We take a number of well known informal definitions of human intelligence that have been given by experts, and extract their essential features. These are then mathematically formalised to produce a general measure of intelligence for arbitrary machines. We believe that this equation formally captures the concept of machine intelligence in the broadest reasonable sense. We then show how this formal definition is related to the theory of universal optimal learning agents. Finally, we survey the many other tests and definitions of intelligence that have been proposed for machines.

Universal Intelligence: A Definition of Machine Intelligence

TL;DR

The paper tackles the fundamental challenge of defining intelligence in a way that applies beyond humans by formalizing a broad, environment-driven notion of machine intelligence. It derives a universal intelligence measure, Υ(π), grounded in a reinforcement-learning framework and weighted across all computable, reward-summable environments using an Occam-inspired prior 2^{-K(μ)} over environment complexity. It then analyzes the properties of this measure, discusses its connection to the AIXI agent, and surveys existing informal and formal definitions/tests of machine intelligence, highlighting strengths, limits, and practicable approximations. The work argues for a theory-grounded, broadly applicable, and potentially testable concept of machine intelligence, while acknowledging computability constraints and the need for practical evaluation protocols. Overall, it provides a rigorous foundation for understanding and comparing intelligent systems across diverse environments and motivates future work on implementable tests that approximate universal intelligence.

Abstract

A fundamental problem in artificial intelligence is that nobody really knows what intelligence is. The problem is especially acute when we need to consider artificial systems which are significantly different to humans. In this paper we approach this problem in the following way: We take a number of well known informal definitions of human intelligence that have been given by experts, and extract their essential features. These are then mathematically formalised to produce a general measure of intelligence for arbitrary machines. We believe that this equation formally captures the concept of machine intelligence in the broadest reasonable sense. We then show how this formal definition is related to the theory of universal optimal learning agents. Finally, we survey the many other tests and definitions of intelligence that have been proposed for machines.

Paper Structure

This paper contains 62 sections, 8 equations, 2 figures, 1 table.

Table of Contents

  1. Introduction
  2. Natural Intelligence
  3. Human intelligence tests
  4. Animal intelligence tests
  5. Desirable properties of an intelligence test
  6. Static vs. dynamic tests
  7. Theories of human intelligence
  8. Ten definitions of human intelligence
  9. More definitions of human intelligence
  10. A Definition of Machine Intelligence
  11. Basic agent-environment framework
  12. Formal agent-environment framework
  13. Agent-environment interaction.
  14. The agent.
  15. Measure of success.
  16. ...and 47 more sections

Figures (2)

  • Figure 1: The agent and the environment interact by sending action, observation and reward signals to each other.
  • Figure 2: A simple game in which the agent climbs a playground slide and slides back down again. A shortsighted agent will always just rest at the bottom of the slide.