On the Computability of AIXI
Authors: Jan Leike, Marcus Hutter
Summary: How might we clear up the machine studying and the unreal intelligence drawback if we had infinite computation? Solomonoff induction and the reinforcement studying agent AIXI are proposed solutions to this query. Each are identified to be incomputable. On this paper, we quantify this utilizing the arithmetical hierarchy, and show higher and corresponding decrease bounds for incomputability. We present that AIXI is just not restrict computable, thus it can’t be approximated utilizing finite computation. Our most important result’s a limit-computable ε-optimal model of AIXI with infinite horizon that maximizes anticipated rewards
