Universal Algorithmic Intelligence

This site is devoted to discussion of “Universal Algorithmic Intelligence” (UAI), which encompasses a family of top-down approaches to learning, decision-making, and ultimately artificial superintelligence (ASI).

The name Universal Algorithmic Intelligence¹ suggests an approach to building and understanding general predictors and agents through algorithmic information theory (AIT), a beautiful subfield of theoretical computer science. Usually, this involves studying predictors and agents that learn rich (world) models represented in the form of probabilistic programs. By unifying learning and decision theory, UAI constructs an elegant, rigorous, and general theory of intelligence.

Programs for knowledge representation have a long history in AI. Programs are a sufficiently rich and flexible ontology to represent a massive variety of possible problem settings – and we speculate this includes interaction with open-ended real world environments. Many practical reasoning systems have been implemented that organize their knowledge as beliefs about generative programs, which model the (probabilistic) rules for generating observations. For instance, computational cognitive scientists are converging on a paradigm called computational rationality which explains cognitive behavior using (automated) inference against probabilistic models in the form of programs. Often, those programs are learned through program induction or program synthesis, well-established subfields of programming language theory. Program induction/synthesis are also sometimes treated as branches of AI. Arguably, UAI is the proper mathematical context for this cross-disciplinary project – in the same way that computational complexity / cryptography are the proper mathematical context for computer security.

A top-down approach means starting from an ideal standard before studying computability, efficiency, and implementation details. The objects studied within UAI are often uncomputable; however, the insights gained from UAI teach us about the limits of what is possible and sometimes guide the search for practical algorithms. Theoretical results from AIT have inspired technologies such as normalized compression distance (NCD) and general RL agents such as MuZero can be considered approximations to AIXI. We hope that UAI will inspire further advances towards safe ASI.

Though our approach is guided by UAI, this community is for people interested in the foundations of super-intelligent agents, very broadly understood. Topics of interest [with reference works] include but are not limited to:

• Algorithmic Information Theory (AIT) and Kolmogorov complexity [Li and Vitanyi]
• Machine Learning (ML) [Bishop]
• Dynamical Systems (DL)
• Sequential Decision Theory (SDT) [Hutter]
• Reinforcement Learning (RL) [Sutton & Barto]
• Artificial Super Intelligence (ASI)
• Artificial General Intelligence (AGI) [Goertzel & Pennachin]
• Universal Inductive Inference (II) and Solomonoff Prediction (SP) [Rathmanner & Hutter]
• Bayesian Statistics (BS) [Press]

More recent research on UAI often focuses on self-reflection and safety properties of universal agents, which seems particularly important in the context of ASI. These topics tie into what is sometimes called “agent foundations.” See the Machine Intelligence Research Institute’s research guide for an overview covering decision theory, Vingean reflection, corrigibility, and other related problems.

See “Getting Started” for a deeper introduction and guide to the research frontier of UAI.