Universal Intelligence: A Definition of Machine 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.

Comparing Humans and AI Agents

“…So, one of the things that we have learnt is that the change of universal distributions from passive environments (as originally proposed in [1] and [3]) to interactive environments (as also suggested in [3] and fully developed in [7,8]) is in the right direction, but it is not the solution yet. It is clear that it allows for a more natural interpretation of the notion of intelligence as performance in a wide range of environments, and it eases the application of tests outside humans and machines (children, apes, etc.…”

“…Later, Legg and Hutter (e.g. [7], [8]) gave a precise definition to the term "Universal Intelligence", also grounded in Kolmogorov complexity and Solomonoff's prediction theory, as a sum (or weighted average) of performances in all the possible RL-like environments. However, in order to make a feasible test by extending from (static) sequences to (dynamic) environments, several issues had to be solved first.…”

“…In the last fifteen years, several alternatives for a general (or universal) intelligence test (or definition) based on Solomonoff's universal distributions [12] (or related ideas such as MML, compression or Kolmogorov complexity) have been appearing on the scene [1,3,7,8,5], claiming that they are able to define or evaluate (machine) intelligence. In this paper we use one of these tests, a prototype based on the anytime intelligence test presented in [5] and the environment class introduced in [4], to evaluate one easily accessible biological system (Homo sapiens) and one off-the-shelf AI system, a popular reinforcement algorithm known as Q-learning [18].…”

On Measuring Social Intelligence: Experiments on Competition and Cooperation

“…Finally, we need to assess the complexity of each task in order to make a proper choice of tasks which capture a wide range of difficulty and, therefore, can suit the agent's level of intelligence. These issues have been addressed in [1,6,2,9,4].…”

“…Dating back from the late nineties, we can find several works [1,2,9,4] addressing the problem of measuring agent intelligence in a principled and general way. Using notions taken from (algorithmic) information theory, MML and twopart compression, Kolmogorov complexity and Solomonoff priors (see [10] for I ?…”