Nanny-Anam Cara interactions example: Conclusion

This scientific approach was applied over 25 years of research on how to model and use context in real-world applications on a spectrum from technology-centred to human-centric applications, that is, from well-defined domains to not formal ones, but all having the goal to model an activity. The presentation is discussed on the example “nanny - anam cara interactions” has all the necessary ingredients to explain the potentiality of the proposed approach.
Our research is part of an approach to designing and implementing AI systems that aim to understand actor(s) through their decisions, actions, and behaviours. Modelling actors’ experience was central to our research and led at a four-level framework: conceptual, operational, implementation and environment levels. For instance, contextual knowledge (conceptual level) is represented as contextual elements (operational level) and designed as a pair of contextual and recombination nodes (implementation level). The model of an activity has two sides, an operational one, on that an actor uses for accomplishing an activity based on a mental model drawn from his mental representation, and an implementation one, a contextual graph that can be used and readable by other actors. The focus of attention for modelling activity allows dividing separation of context in contextual knowledge and external knowledge. The explicit integration of context in the representation (through contextual elements and their instantiations) follows the human style of actors’ activity (collecting and structuring information, making decisions, and acting). On the AI side, the CxG formalism of representation plays the role of a "concept revealer" in a model.
We consider that a mental model is either a path in the contextual graph (in actor activity modelling) or a sequence of independent subtasks that define actors’ activities (in group activity modelling). The mental model is developed from the mental representation in the actor version, but initially must be built in real time from independent subtasks and then developed in the group version. The changes in the group version, with respect to the actor version, are the recording of independent subtasks in the mental representation instead of mental models and the cyclic use of the contextual graph to build a mental model. The notion of group activity is dynamically modelled at two levels: first, at an operational level (turn sequences), and second, at the implementation level (cyclic use of the directed contextual graph). Another important concept is the shared context that makes possible the cyclic use of a directed, acyclic and series-parallel contextual graph and the existence of CxG-based simulation as a natural function of the CxG software. The shared context is used as an inference engine for group-activity building, the engine assuring the turn mechanism in CxG-based simulation. A turn is a local contribution of an actor to the group activity, and the turn mechanism plays a synchronizer role in the dynamic assembling of independent subtasks for building mental models, thanks to reserved contextual elements that monitor turn management. The CxG-based simulation is a function of the CxG formalism for group activity. This tool also offers the possibility of managing other tasks simultaneously (jointly with their realization), such as negotiation, changes in objectives, and looking ahead, thanks to context management. It is possible to “replay” the simulation in different contexts.
Contextual reasoning explains the mental-model development as a path from the input to the exit of the contextual graph, on which contextual elements are instantiated. Contextual reasoning can be nonlinear (e.g. g., local search, voting system, or the Contextualisation-Decontextualization-Recontextualization approach) (Brézillon 2023), and contextual elements themselves, with their implementation as pairs of contextual and recombination nodes, behave as units of contextual reasoning at an operational level. The CxG formalism is effective for modelling an activity, not for visualising its evolution. A tree representation supports a simple visualisation of contextual reasoning (and all its known variants) in the CxG formalism. The mental-model tree view shows to actors the relevant contextual elements as a proceduralized context (the ordered sequence of instantiated contextual elements) and postpones actions to quickly make decisions.
By putting context front stage in the Contextual-Graphs formalism, we obtain a uniform representation of knowledge, information, reasoning and context coming from sources of different natures. We thus have been able to model activities in very different domains (subway, army, different types of cancer in medicine and workflows), thanks to the Contextual-Graphs formalism that is very simple to use. Finally, the CxG formalism is a passport for intelligent systems based on human experience. The “hard kernel” of our approach is the explicit modelling of context in activity, which leads to a homogeneous view of how a class of AI systems can become context-based intelligent systems, especially context-based intelligent assistant systems (CIASs) (Brézillon 2023) which aim at reuse and extend human experience based on how this experience grows. CIASs developed in the CxG formalism offer the possibility to model contextual reasoning with context-based simulation, a powerful modelling tool for CIASs.