Model-based preference quantification

0573588 - ÚTIA 2024 RIV GB eng J - Článek v odborném periodiku
Kárný, Miroslav - Siváková, Tereza
Model-based preference quantification.
Automatica. Roč. 156, č. 1 (2023), č. článku 111185. ISSN 0005-1098. E-ISSN 1873-2836
Grant ostatní: EU-COST(XE) CA21169
Institucionální podpora: RVO:67985556
Klíčová slova: Dynamic performance * Probabilistic * Preferences * Optimal strategy * Preference elicitation * Exploration
Obor OECD: Automation and control systems
Impakt faktor: 6.4, rok: 2022
Způsob publikování: Omezený přístup
http://library.utia.cas.cz/separaty/2023/AS/karny-0573588-preprint.pdf https://www.sciencedirect.com/science/article/pii/S0005109823003461?via%3Dihub

Any prescriptive theory of decision-making (DM) has to cope with the common DM agents’ inability to fully specify their preferences dependent on several attributes. The paper provides the needed preference completion and quantification for fully probabilistic design (FPD) of DM strategies. FPD (covering the usual Bayesian DM) probabilistically models the agent’s environment and quantifies its preferences via an ideal probabilistic model of the closed DM loop. The probability density (pd) models (closed-loop) behaviour, a collection of involved random variables. Its ideal twin is high on desired behaviours, small on undesired and zero on forbidden ones. The FPD-optimal strategy minimises the Kullback-Leibler divergence (KLD) of the closed-loop modelling pd to the ideal twin. The exposed preference quantification chooses the optimal ideal pd from the set of pds compatible with partially-specified agent’s preferences. The optimal ideal pd minimises the KLD minima reached by the optimal strategies for respective imminent ideal pds. This preference-focused twin of the minimum KLD principle was applied to special sets of ideal pds. The paper extends them towards exploration and balancing contradictory wishes on states and actions.
Trvalý link: https://hdl.handle.net/11104/0344021