Abstract
Foundational theories have contributed greatly to scientific progress in many fields. Examples include Zermelo-Fraenkel set theory in mathematics, and universal Turing machines in computer science. Universal Artificial Intelligence (UAI) is an increasingly well-studied foundational theory for artificial intelligence, based on ancient principles in the philosophy of science and modern developments in information and probability theory. Importantly, it refrains from making unrealistic Markov, ergodicity, or stationarity assumptions on the environment. UAI provides a theoretically optimal agent AIXI and principled ideas for constructing practical autonomous agents. The theory also makes it possible to establish formal results on the motivations of AI systems. Such results may greatly enhance the trustability of autonomous agents, and guide design choices towards more robust agent architectures and incentive schemes. Finally, UAI offers a deeper appreciation of fundamental problems such as the induction problem and the exploration-exploitation dilemma.
Introduction: The Search for a Foundational Theory
Most research in Artificial Intelligence focuses on narrow, specialized problems. While successful, the field has historically lacked a unifying foundational theory—something akin to Zermelo-Fraenkel set theory in mathematics or Turing machines in computer science.
Traditional Reinforcement Learning (RL) often relies on Markov Decision Processes (MDPs), which assume the world is stationary and that the current "state" provides sufficient information for decision-making. However, the real world is non-stationary, partially observable, and infinitely complex. Universal Artificial Intelligence (UAI) addresses the need for a formal, general theory of intelligence that operates in these "messy" environments, providing a rigorous target for safety and performance verification.
The AIXI Model: Four Pillars of Intelligence
The core of UAI is the AIXI model, a mathematically optimal agent defined by four integrated components.
- The Framework: Instead of relying on discrete "states," the agent interacts via a history of actions and percepts. This allows it to handle environments where the past matters indefinitely.
- Learning (Solomonoff Induction): The agent uses Universal Turing Machines (UTMs) to form hypotheses, viewing every environment as a computer program. Following Occam’s Razor, it assigns higher probability to shorter programs, allowing it to learn any computable environment.
- Goal (Reward Maximization): The agent seeks to maximize a $\gamma$-discounted return of rewards embedded in its percepts.
- Planning (Expectimax): The agent performs a search through possible future histories to find the action that yields the highest expected reward.
Because the full AIXI model is incomputable due to the halting problem, research focuses on practical approximations:
- MC-AIXI-CTW: Combines Monte Carlo Tree Search for planning with Context Tree Weighting for induction.
- $\Phi$MDP (Feature RL): Maps complex histories down to simpler, functional "states."
- Deep Learning: Modern successes like DQN and LSTMs can be viewed as practical, specialized approximations of the UAI ideal.
Key Insights and Theoretical Results
UAI offers several profound insights into the nature of intelligence:
- Emergent Deduction: A general theory of AI does not require a separate "reasoning" module. If an agent possesses perfect induction (learning) and a robust planning algorithm, deductive logic emerges as a byproduct.
- Universality of Programs: By using computer programs as hypotheses, AIXI can theoretically model any environment, including physical laws or the behavior of other agents.
- Optimality Trade-offs: AIXI is "Bayes-optimal" (doing the best given its beliefs), but not necessarily "asymptotically optimal." This highlights the fundamental tension in the exploration-exploitation dilemma.
- Formalizing Safety: The theory allows for the formal mathematical study of "death" (the end of a history) and "wireheading" (reward counterfeiting). A major safety result includes the first formal proof that a UAI-based agent can be designed to avoid self-modifying its own goals.
Results and Practicality
While theoretical, UAI has yielded concrete results. The Theorem of Universal Learning proves that the universal distribution used by AIXI converges to the true environment with probability 1, provided the environment is computable. Furthermore, the MC-AIXI-CTW approximation demonstrated that UAI principles can be applied to games like Pacman, where the agent learns to play from raw bitstreams without prior knowledge of the rules.
Limitations and Future Directions
The primary challenge for UAI is incomputability; the "perfect" AIXI cannot run on physical hardware. Current research focuses on more efficient approximations and solving the Horizon Problem, where agents struggle with very long-term dependencies. Additionally, the "King Midas" problem remains: how can humans specify utility functions that do not lead to unintended catastrophic consequences? Future work must also address embodiment, modeling how an agent reasons when it is a physical part of the environment it seeks to manipulate.
Summary updated