Detecting Outlier Intelligence in the behavior of intelligent coalitions of agents

Abstract

In our research, we consider the measuring of machine intelligence based on the intelligence (ability to solve various tasks in high efficiency, with a grade of flexibility and robustness) in solving difficult problems/tasks (NP-hard and/or have different types of uncertainties). An intelligent cooperative coalition of agents (could be a whole cooperative multiagent system or a part of it) have variability in the problem-solving intelligence. It is able to solve more or less intelligently different problems. For some problems solving it could manifest even extremely low or extremely high intelligence. We call such intelligence values as outlier intelligence, which could be low outlier intelligence values or high outlier intelligence values. In this paper, we propose a novel method called OutIntDet (Outlier Intelligence Detection Method) for detecting outlier intelligence values. In order to sustain the effectiveness of the proposed method, a case study where we considered a coalition of agents which solve an NP-hard problem was performed. OutIntDet could be useful to be implemented in some intelligence metrics that are based on measuring problems-solving intelligence, being able to detect low and high outlier intelligence. This could make the metrics more accurate and robust. OutIntDet is also appropriate for the identification of the problems for whose solving the coalition manifest very low or very high intelligence. There is also presented an appropriate calculation of the MIQ (Machine Intelligence Quotient) based on the properties of measured problem-solving intelligence data.