A framework for establishing shared, task-oriented understanding in hybrid open multi-agent systems

In Open Multi-Agent Systems (OMAS), the open nature of such systems precludes that all communication protocols are hardwired in advance. It is therefore essential that agents can incrementally learn to understand each other. Ideally, this is done with a minimal number of a priori assumptions, in order not to compromise the open nature of the system. This challenge becomes even harder for hybrid (human-artificial agent) populations. In such a hybrid setting, the challenge of learning to communicate is exacerbated by the requirement to do this in a minimal number of interactions with the humans involved. The difficulty arises from the conflict between making a minimal number of assumptions while also minimizing the number of interactions required. This study provides a fine-grained analysis of the process of establishing a shared task-oriented understanding for OMAS, with a particular focus on hybrid populations, i.e., containing both human and artificial agents. We present a framework that describes this process of reaching a shared task-oriented understanding. Our framework defines components that reflect decisions the agent designer needs to make, and we show how these components are affected when the agent population includes humans, i.e., when moving to a hybrid setting. The contribution of this paper is not to define yet another method for agents that learn to communicate. Instead, our goal is to provide a framework to assist researchers in designing agents that need to interact with humans in unforeseen scenarios. We validate our framework by showing that it provides a uniform way to analyze a diverse set of existing approaches from the literature for establishing shared understanding between agents. Our analysis reveals limitations of these existing approaches if they were to be applied in hybrid populations, and suggests how these can be resolved.