To develop intelligent computational systems that incorporate human behaviour, models of human processes are an important building block. In our research, we have developed models for both individual processes and multi-agent (social) processes. Moreover, mutual (abstraction) relationships between different models have been investigated. Traditionally, knowledge is exploited from the areas of Cognitive Science and Social Sciences to design single agent models for human processes,. This knowledge does not always provide a biologically plausible basis for underlying mechanisms. However, in recent years much progress has been made in the areas of Cognitive, Affective and Social Neurosciences in revealing the underlying biological causal mechanisms. This type of knowledge about causal mechanisms is a useful source that is exploited more and more in this area of research, as a basis for human-like agent models. A number of specific human processes have been addressed from a dynamical systems perspective, using the hybrid temporal-causal modelling language LEADSTO or using difference/differential equations. These models can be used for the development of realistically behaving virtual agents, for example for in simulation-based training, or for human-aware systems.
In the case of the single agent model, the following has been addressed from the affective side: emotion and feelings and their integration in cognitive processes, emotion and mood regulation; desiring, feeling and temptation, depression and the effect of therapies, fear extinction learning, and posttraumatic stress disorders. Moreover, attention, situation awareness, and functional state (e.g., exhaustion and fatigue) have been addressed. Furthermore, decision making has been investigated (experiences, rationality and emotion-related valuing; psychological and practical factors influencing behaviour change; simulated behaviour and perception chains; biological vs rational factors), trust, surprise, greed and risk behaviour and habitual behaviour. Finally, adaptivity and plasticity within many of these models have been studied.
For the multi-agent case, models have been developed for humans in a social context. A central concept exploited from Social Neuroscience is mirroring, which is the basis for mutual influences from one person to the other by contagion. Amongst the issues addressed are; emotion reading, mindreading, theory of mind, empathy and natural variation in social functioning. In a wider context, social contagion (of emotions, intentions, beliefs, and integration), social support networks, team development, collective decision making and crowd behaviour have been addressed.
Computational models for human processes play a crucial role in this area of research . However, such models can be made at different abstraction levels, and it can be analysed how models at different levels relate to each other. The purpose of the model defines which abstraction level is most adequate. For example; knowledge of mechanisms provided by neuroscience is a useful source for modelling. However, it is more efficient to abstract these mechanisms to causal relationships modelled at a cognitive and affective level, than to model them neuron-by-neuron. Moreover, it may even be considered to abstract such a model to a behavioural model, abstracting from any internal cognitive or affective state. Usually the heuristic is followed that a model should not be made at too low alevel of abstraction for the accuracy required, as this entails unnecessary complexity. However, how far a deviation is introduced by doing so is a serious question that needs analysis. Therefore in this research area, models at different abstraction levels and their (interlevel) relations are being investigated. Three dimensions of abstraction covered are: process abstraction (neurological/physiological vs cognitive/affective vs behavioural), agent clustering (single agent-based vs population-based, group abstraction), and temporal abstraction (causal models for underlying mechanisms vs emerging temporally global properties). For each of these dimensions a number of results have been achieved. These results address: relations between cognitive/affective models and neurological models, and between behavioural models and cognitive/affective models; relations between agent-based models and population-based models; relations between causal models of mechanisms and emerging temporally global properties.
Intelligent systems that take human processes into account should be able to maintain a form of awareness of human processes and states over time: human-aware systems. To obtain such human-awareness, monitoring based on sensoring can be applied, as usually advocated, for example, in Ambient Intelligence. However, most human processes and states cannot be monitored by sensoring alone, at least not directly. Nevertheless, with the help of human-like computational dynamic models that describe how the processes that cannot be sensed relate to other states for which sensing is possible, it might be possible to monitor those. It has been explored how by integrating such human-like models, a human-aware intelligent system can be obtained, in which the human-awareness extends much beyond what is possible only with direct monitoring. The human-like models can be used both for analysing and assessing the human processes and states, and determining adequate support actions based on such assessments. This is achieved by integrating model-based reasoning methods, both for qualitative and quantitative dynamical models. Moreover, it has been investigated how such processes can be made more efficient by using more abstracted types of human-like models, as a form of compilation (e.g., based on representational content specifications for internal states, and by compiling heuristic rules for support actions). Furthermore, it has been investigated how parameters in such human-like models (which usually represent person-specific characteristics) can be tuned by automated methods.
The types of human-like and human-aware agent models investigated in the Social AI research area aim at a number of application contexts for which obtaining a successful application is important. The following application contexts have been addressed.
To support individual persons, two main types of application contexts have been addressed. The first type can be characterised as support for professionals in demanding circumstances. As an example, in cooperation with the Royal Netherlands Navy, this has addressed human-aware agents supporting naval officers with respect to their attention, functional state, workflow, and team emotions. A second type of application context, which shows a kind of opposite situation but with similar discrepancy, is support for demanding persons in ordinary circumstances. Here 'demanding persons' is an abstracted term for people needing enhanced efforts to function in daily circumstances. For example, this has been addressed for people with chronic diseases (e.g., diabetes, HIV), for people with depression, and those vulnerable to temptations (e.g., addictive or criminal activities).
AAnother application context addressed, is the use of (human-like) virtual agents, for example; for serious gaming in simulation-based training applications or to represent or evaluate persons' states in order to make them aware of these states. Such applications have been addressed concerning emotions in virtual agents, virtual agents to illustrate the role of greed and risk in financial decision making, for fighter pilot training, virtual soccer players, virtual opponents in entertainments games, and virtual agents to prevent cyberbullying.
In a social context, support may go beyond the individual case. Human-aware agents have been investigated that support a person's social network, for example, informal caregivers or friends of persons with depression. In these cases the human-aware agent does not only use human-like models for individual functioning, but also for social interactions within groups or networks.
Finally, the application area of social and socio-technical systems has been addressed at different levels. This concerns multi-agent-based design and analysis methods for social and socio-technical systems and organisations (small-scale and large-scale). Such systems have been addressed, for example, in the analysis of Air Traffic Management and safety, social emergency situations (e.g., emergency evacuations), analysis of crime prevention policies, and predicting development of juvenile delinquency.