Interacting Decision-making Agents and their Impacts on Assurances: Taxonomy and Challenges

[Context/ Motivation] Decision-making requires the quantification and trade-off of multiple software quality assurances (aka non-functional requirements or softgoals) and the analysis of benefits and costs between alternative functional goal realizations. Modern self-adaptive and self-organizing systems usually must achieve several functional goals. Each can be achieved through alternative realization strategies (i.e. choices) and supported by the decision process of one agent or subsystem. The above is an example, among many others, that implies considering a system shaped by several agents interacting among themselves in a decentralized way towards a higher level goal. [Problems/Questions] The multiple interacting agents involve diverse perspectives with effects on different common software quality assurances while the agents collaborate or compete among themselves. An example of a potential problem could be that one decision suggested by an agent may cancel out the positive effects of another agent on a given software quality with potentially unwanted results. The agents can also use machine learning techniques which exacerbates the uncertainty implied and its effect on the interactions among agents over time. We argue that there is the need for coordinating the decision-making process by different agent-based models in an explicit and decentralized way and that models@run.time can be a means for this. The runtime coordination gives place to big challenges. [Principal ideas] In this paper, and based on an initial exploration of the state-of-the-art, we discuss ideas about how to engineer modern self-adaptive and self-organizing systems. We propose a taxonomy to ease the engineering of such systems and based on that, we discuss the role of models@run.time to support the models of expected behaviour in an explicit way.