2015/11/05 by Jordi Grau-Moya, Grau-Moya, Jordi, Daniel A. Braun +1
Computer Science · Decision Sciences · Engineering · Mathematics · Social Sciences · #Algorithm #Artificial Intelligence (cs.AI) #Artificial intelligence #Bayesian probability #Bounded function #Bounded rationality #Computer science #Data compression #Decision-Making and Behavioral Economics #Entropy (arrow of time) #Experimental Behavioral Economics Studies #FOS: Computer and information sciences #Heuristics #Mathematical optimization #Mathematics #Partition (number theory) #Prior probability #Rate–distortion theory #Theoretical computer science #Water resources management and optimization #cs.AI
paper · pdf · doi:10.48550/arxiv.1511.01710
published in arXiv (Cornell University), 1-4 (Cornell University) · 4 pages, 1 figure, Workshop on Bounded Optimality and Rational Metareasoning at Neural Information Processing Systems conference, Montreal, Canada, 2015
arxiv created 2015/11/05 · openalex publication_date 2015/11/05 · arxiv updated 2015/11/06 · openalex created_date 2022/09/02 · openalex updated_date 2026/07/28
Deviations from rational decision-making due to limited computational\nresources have been studied in the field of bounded rationality, originally\nproposed by Herbert Simon. There have been a number of different approaches to\nmodel bounded rationality ranging from optimality principles to heuristics.\nHere we take an information-theoretic approach to bounded rationality, where\ninformation-processing costs are measured by the relative entropy between a\nposterior decision strategy and a given fixed prior strategy. In the case of\nmultiple environments, it can be shown that there is an optimal prior rendering\nthe bounded rationality problem equivalent to the rate distortion problem for\nlossy compression in information theory. Accordingly, the optimal prior and\nposterior strategies can be computed by the well-known Blahut-Arimoto algorithm\nwhich requires the computation of partition sums over all possible outcomes and\ncannot be applied straightforwardly to continuous problems. Here we derive a\nsampling-based alternative update rule for the adaptation of prior behaviors of\ndecision-makers and we show convergence to the optimal prior predicted by rate\ndistortion theory. Importantly, the update rule avoids typical infeasible\noperations such as the computation of partition sums. We show in simulations a\nproof of concept for discrete action and environment domains. This approach is\nnot only interesting as a generic computational method, but might also provide\na more realistic model of human decision-making processes occurring on a fast\nand a slow time scale.\n