Intelligent Agent Control and Coordination with User-Configurable Key Performance Indicators

Abstract

In Multi-Agent Systems (MASs) for autonomous control of industrial and logistic processes, intelligent agents are often faced with multi-criteria decision problems. The agents’ local goal systems may comprise qualitative and quantitative objectives that can conflict at the individual agent’s local level as well as on a global scope between agents. In general, the problem to be solved by the MAS cannot be decomposed in a way that eliminates all these conflicts as the competing goals are an integral property of the problem itself. Consequently, acceptable trade-offs need to be identified and negotiated by the agents. Due to the emergent, not entirely predictable behavior of many complex MASs, different users of the system may have different views and requirements regarding what constitutes an acceptable trade-off, or they may need to experiment with different goal settings before the system starts exhibiting the desired behavior. In this paper the concept of numerical key performance indicators is applied to agent control and coordination in MASs. A software framework is presented which allows the user of a MAS to define at runtime numerical key performance indicators and quantitative objectives attached to them, which then are incorporated into the agents’ individual goal systems in order to influence the local as well as global agent behavior. The central parts of the framework have been implemented as a Java programming library that facilitates the assessment and optimization of key performance indicators in MASs.