Realizing the Potential of Agent-Based Social Simulations

Computer simulation has been established as a method for investigating the interplay of mechanisms, complex phenomena, or the general behavior of a system under certain conditions. When real-world experiments are too costly, time-consuming or impracticable, simulation provides an alternative approach for controlled and systematic investigations. Experiments are executed with artificial systems but allow for drawing conclusions about the real-world system, i.e., for deriving insights about its functioning based on observations that were made. Such insights cannot only be used by scientists to formulate new theories, but maybe more importantly by policy- and decision-makers to consider the advantages and disadvantages of possible measures, rules, or actions before they are implemented.

There are many different simulation methods and paradigms, but typically they have problems to model complex human behavior and social interaction in a realistic way. As an attempt to address these issues, the paradigm of Agent-Based Social Simulation (ABSS) has been introduced. ABSS is a very powerful simulation paradigm that integrates AI in the modelling of human behavior and social interaction. The relevant properties of each individual, as well as its decision-making and actions, are explicitly modelled. However, although there are a few successful applications of ABSS that have informed actual policy- and decision-making, e.g., for electricity markets or emergency management, the majority are from research projects providing useful insights to social scientists.

We argue that the full potential of ABSS yet has not been realized in terms of providing support for societal decision-makers. One problem has been the scalability of ABSS, which often works great when the number of individuals is in the thousands but not in the millions. Moreover, the models of human behaviour used in current ABSS are rather homogenous, often not taking into account the actual variations in populations, with respect to e.g. age, gender, ethnicity, cultural characteristics and habits. In order to provide support for decision-making in a responsible way that is ethically sound, fair and inclusive, the modelling of this diversity is crucial. Furthermore, policy- and decision-makers and their analysts have experience from and trust in the traditional simulation paradigms, and hesitate to use ABSS. Another reason for this hesitation is the amount of input parameters typically needed in ABSS, which however could also be viewed as one of its strength. It requires the modeler to be more explicit about the assumptions she/he makes. A current application area is the simulation of pandemics and the effects of counter measures. Some simulation models just have a single parameter, e.g. the reproduction number, which is supposed to capture all of the individuals’ behavior and interaction. Thus, assumptions about the effects of social distancing, mobility, work at home, etc., are hidden in this parameter. By using ABSS, the assumptions about all these aspects could be made explicit and significantly affect the results.

Due to the variety of existing simulation models, it is not only challenging for decision-makers to find appropriate models, but also to assess individual strengths and weaknesses. Models differ for instance in the selected simulation paradigm (e.g., system dynamics or individual-based), perspective (micro, macro), considered scenarios, assumptions they make, output metrics they provide, scale, and the framework or language they use. For analysts, this oversupply in similar models can lead to insecurity on which model to use and trust. To overcome this issue, one approach is to combine several models to multi-model ensembles to provide better results and reduce uncertainty compared to each individual model. This is an established approach in e.g. machine learning and weather forecasting, mainly due to error diversity, assuming that model errors will cancel each other out. There is a lack of similar approaches to combine simulation model of human behavior, to allow for more comprehensive investigations, and to overcome individual weaknesses. The idea of using ensembles of agent-based models has been suggested, but research is still in its infancy. It requires the identification of suitable AI methods for the construction of ensembles, proper quality assessment, and interfaces for the use of ensembles by researchers and practitioners across different disciplines.