The «k-consensus» decision-making rule for a group of artificial intelligent actors at the second stage of the evolutionary matching method

Abstract

The article presents a decision‑making rule for a group of artificial intelligent actors called «k‑consensus», designed for the second stage of the evolutionary matching method – the stage of evaluating actors' responses. The developed rule makes it possible to analyze the responses of several machine learning classification models and, based on this analysis, to form a group decision that, according to Condorcet’s theorem, has an advantage over individual decisions in terms of accuracy and relevance. In addition to positive and negative decisions, the developed rule allows the group of actors to express doubt and respond with “I don’t know”, which implies the transfer of the right to make a decision to a human. This response is ensured by evaluating the decisions of the actors using the root‑mean‑square deviation, Shannon entropy, and the general average of the confidence metrics for each actor. This rule has a limitation – it can be applied only to non-composite decisions of actors. Formulas have been derived to calculate the values of the boundaries of the group’s uncertainty zone, and a necessary and sufficient condition for the existence of this zone has been presented. Experimental testing of the developed rule has shown that it significantly outperforms classical voting rules in cases where it is permissible to refrain from making a decision in disputed situations. A study of the rule’s sensitivity to changes in parameters was conducted, which demonstrated its stability and controllability.