"Approximately Universally Intelligent" — How AIXI can Inspire Practical AI Algorithms

Paper

Universal Artificial Intelligence (UAI) is a mathematical framework that produces a provably optimal and maximally general agent called AIXI, serving as a theoretical gold standard for artificial general intelligence. However, AIXI is inherently incomputable and thus cannot be implemented directly in practice. Under the lens of generality, this report discusses the merits and limitations of UAI as well as practicable approximations thereof. Overall, UAI provides strong theoretical optimality guarantees and flexible framework modularity that allows adopters to, in principle, “choose” the level of generality. However, the framework makes potentially generality-narrowing modelling assumptions and the few existing AIXI approximations lack quality guarantees with respect to generality. UAI thus remains mainly a theoretical subject for now, with a myriad of open research avenues to explore before it becomes evident whether UAI is the ultimate path to AGI.