Output details
11 - Computer Science and Informatics
University College London
A Monte-Carlo AIXI Approximation
<12>This was the first paper to achieve a practical and high-performance approximation to a universally optimal agent. It builds on Hutter's AIXI: a universal, optimal agent for artificial intelligence - which is, unfortunately, incomputable. The agent developed in this paper is both formally proven to converge to universally optimal decisions (within a restricted class of k-Markov environments); was empirically shown to outperform previous approaches; and was validated on significantly harder problems than previous approaches. Our AIXI approximation is now used commercially by DeepMind Technologies to build general-purpose agents in various domains. Journal version of the 2010 AAAI conference paper.