Mathematically defining intelligence is not only possible, it is also crucial to understanding and developing super-intelligent machines. The equation that we have developed is below and formalises the informal definition of intelligence, namely an agent’s ability to succeed or achieve goals in a wide range of environments.

In a nutshell, AIXI has a planning component and a learning component. Imagine a robot that initially has no knowledge about the world. It acquires information from the world from its sensors and constructs an approximate model of how the world works. New observations allow AIXI to improve its world model. AIXI uses this model to predict future events and bases its decisions on these tentative forecasts. The goal of AIXI is to maximise its reward over its lifetime.

Since AIXI is incomputable, it needs to be simplified in practice. Our work is in the early experimental phase: the approximation can learn to play Pac-Man, Tic-Tac-Toe and other games from scratch. Ultimately, we are developing rigorous foundations for intelligent agent systems, which are prerequisites for the creation of more flexible, adaptive, robust, reliable and secure software that our modern society needs.

Our interdisciplinary research on intelligent agents involves machine learning, reinforcement learning, artificial intelligence, information theory and statistics. AIXI integrates numerous philosophical, computational and statistical principles including Ockham’s razor, Bayes rule and Bellman equations