**7. User Behaviour Modelling**

Modelling the adaptive user and system behaviour in dynamic non-sequential interactions is a core advantage of the SemMR framework. This is of key importance for a number of settings, including (a) cultural sites and cities with tourist attractions, since their visitors are free to roam and view the site points of interest in no particular order and (b) learning environments, where learners are able to choose between learning paths or access and use learning material in distinct sequences, collaborating with other learners or instructors. Towards this, cognitive models that produce detailed simulations of human (multi-) task performance were designed and used to implement simulated artificial agents in a multi-agent (multi-entity) setting. AI agents compute the most plausible task action(s), given their understanding of the context, actions of others, their preferences and goals, provide alternatives and plans, roll out possible outcomes, and, therefore, are able to adapt their behaviour to their partners. They also know why they select a certain action and can explain why the choices made lead to each specific outcome (explainable AI). Agents can be built while using rather limited real or simulated expert and/or interactive data: an agent is supplied with initial state-action templates encoding domain knowledge (as eX-trajectories enriched with VE and IoT information), the user's profile, and preferences. Over time, the agent learns from the collected interactive experiences. Suggestions and optimizations are performed by finding prior experiences (instances) that are the most frequently or most recently used and/or are most similar to the current situation (contextual parameters, user affective state, user's goals, and preferences), blending the instances together to the extent that they match the interactive state.

The advantage of the SemMR approach is that it requires far less experiences for the system to be able to interact with the user in a sensible way and that it incrementally improves as its set of instances increases in size. It also allows for utilizing experiences of others to guide and enrich the experiences for new users. Additionally, growing data from the eX-trajectories store, paired with increased computational power, can be used to apply modern powerful Artificial Neural Networks (ANNs) and Deep Learning (DL) approaches, which showed a significant impact on many AI and HCI applications [61,62]. Finally, the SemMR model accounts for real-time user interaction errors, which supports the modelling of dialogue repair in user understanding, and the integration of a memory and interest model, reflecting the individual and changing configurations of the user's mind.
