A Universal Measure of Intelligence for Artificial Agents

Abstract

A fundamental difficulty in artificial intelligence is that nobody really knows what intelligence is, especially for systems with senses, environments, motivations and cognitive capacities which are very different to our own. If we look to definitions of human intelligence given by experts, we see that although there is no consensus, most views cluster around a few common perspectives and share many key features. In all cases, intelligence is a property of an entity, which we will call the agent, that interacts with an external problem or situation, which we will call the environment. An agent’s intelligence is typically related to its ability to succeed in environments, which implies that there is some kind of objective, which we will call the goal. The emphasis on learning, adaptation and flexibility common to many definitions implies that the environment is not fully known to the agent. Thus intelligence is the ability to deal with a wide range of possibilities, not just a few specific situations. Putting these things together gives us our informal definition of intelligence: Intelligence measures an agent’s general ability to achieve goals in a wide range of environments. We are confident that this definition captures the essence of many common perspectives on intelligence. It also describes what we want to achieve in machines: A very general capacity to adapt and perform well in a wide range of situations. In this paper we will try to formalise this view of intelligence.