Stability of artificial intelligence controlled systems

Abstract

The impact of artificial intelligence in novel design of system of systems is substantially increasing, having influenced overall behavioral pattern of full setup. Major impact could be seen in large systems equipped even with minimum artificial intelligence elements up to specifically designed systems e.g. for evaluation of contribution on social network. The basic characteristic of any artificial intelligence program is that creator cannot predict next steps in the behavior of the whole system controlled by artificial intelligence. It can be done only based on statistical estimation, because real artificial intelligence program cannot behave deterministically. Moreover, the artificial intelligence programs today are not capable to demonstrate such a quality as intuition or deep abstraction etc., which are specific to human. The concentration of the cyber systems as are sensor and actuators network in IoT environment or autonomous car groups under control of artificial intelligence of the similar design can be source of instability of the system itself.When we talk about a system stability, we usually consider the system with some observable feedback and short control loops. In the large system of systems, the control loops are multiple, hidden and mutually affected. In our contribution we are going to show the possible way of analysis of such a case of dynamic system behavior based of semidefinite programming approach for both linear and non-linear system with and without feedback loop. Naturally, where the feedback loop will be missing, the stability of the system will be very sensible to border conditions while in the case of very tight loop the system will be kept to one specific state. Both of these marginal solutions are not acceptable for large systems equipped with artificial intelligence.The goal of system of systems equipped with artificial intelligence would consist in ability to cover very large number of possible states, often vaguely defined, based on probability of the border conditions. The system of systems remains in the state achieved, analyzing its state itself and implementing change into another state under reasonable change in the environment. So it means system of systems will remain in the probabilistically optimal state until the self-analysis will show necessity to change this state to another, more optimal, state. But change in the state can lead to instability of the system of systems when probabilistic estimation will fail or environmental condition are not in the form known from the past. Based on this analysis we expect to show, that system of systems, where the internal systems are using artificial intelligence, could be statistically stable only. Hereto such a system could create a threat, when the coupling among the individual systems will be weak leading to disbalance of all system or vice versa when the system is strongly coupled lead by one predominant artificial intelligence and in such a case showing the groupthinking effect.