Bibliography
Conference Paper (international conference)
Experiments with the User’s Feedback in Preference Elicitation
,
: AIABI-2022 : Artificial Intelligence and Applications for Business and Industries 2022
: Artificial Intelligence and Applications for Business and Industries 2022, (Udine, IT, 20221127)
: Preference elicitation, Adaptive agent, Decision making, Bayes rule
: http://library.utia.cas.cz/separaty/2023/AS/sivakova-0575198.pdf
(eng): This paper deals with user’s preferences (wishes). Common users are uneducated in the decision-making (DM) theory and present their preferences incompletely. That is why we elicit them from such a user during the DM. The paper works with the DM theory called fully probabilistic design (FPD). FPD models closed DM loop, made by the user and the system, by the joint probability density (pd, real pd). A joint ideal pd quantifies the user’s preferences. It assigns high probability values to preferred closed-loop behaviors and low values to undesired behaviors. The real pd should be kept near the ideal pd. By minimizing the Kullback-Leibler divergence of the real and ideal pds, the optimal decision policy is found. The presented algorithmic quantification of preferences provides ambitious but potentially reachable DM aims. It suppresses demands on tuning preference-expressing parameters. The considered ideal pd assigns high probabilities to desired (ideal) sets of states and actions. The parameters of the ideal pd (tuned during the DM via the user’s feedback) are: ▶ relative significance of respective probabilities. ▶ a parameter balancing exploration with exploitation. Their systematic tuning solves meta-DM level task, which observes the agent’s satisfaction expressed humanly by “school-marks”. It opts free parameters to reach the best marks. A formalization and solution of this meta-task were recently done, but experience with it is limited. This paper recalls the theory and provides representative samples of extensive up to now missing simulations.\n
: BC
: 10201