Large-scale agent-based social simulation: A study on epidemic prediction and control

Large-scale agent-based social simulation is gradually proving to be a versatile methodological approach for studying human societies, which could make contributions from policy making in social science, to distributed artificial intelligence and agent technology in computer science, and to theory and modeling practice in computer simulation systems. Simultaneously, the application areas of largescale agent-based social simulation vary a lot as well, from daily transportation on a city/country level, to large-scale emergency response, prediction of social change, and analysis of social structure.
However, current large-scale agent-based social simulation practice is facing difficulties in balancing model complexity and simulation performance. The wide adoption of distributed/parallel mechanism in current large-scale agent-based social simulation has proven to be an efficient solution to achieve system performance and scalability. On the other hand, the trade-off is usually the simplification of the model precision including agent behavior, agent environment and the social networks and interactions, which are proven to be important to understand social phenomena in complex social systems.
Based on the existing challenges, this thesis introduces a novel conceptual model for large-scale agent-based social simulation development, gives out the reference implementation of the proposed model components, and presents a simulation study of a case of epidemic prediction and control in the city of Beijing. This conceptual model can be considered as a hybrid model mixing a general agent-based conceptual model and the discrete event simulation paradigm.
For the concept of agent in the proposed conceptual model, this thesis presents a new way for implementation. A reference implementation of an agent is constituted by three main parts: (1) agent object, (2) activity pattern, and (3) multi-level decision-making module. With this design, the implemented agents can carry out a lot of complex activities and show diverse behaviors, such as traveling around and joining non-predefined social activities, while staying "simple" and "small" enough for scalability consideration.
For the concept of a social network, this thesis presents a new method to generate social networks dynamically for simulating interactions among a group of agents on a large scale during a simulation run. This thesis borrows from the concept of ’social reach’ in a social circle model, and proposes the concept of ’social similarity’ to generate the special type of social networks, friendship. Using the generated entire social network, agents in this model are able to communicate for scheduling joint social activities. When executing joint social activities, a functional entity called ’activity group’ is generated to organize and manage the participants, and a social contact network emerges from the execution.
Compared to the concept of agent environments in general ABM conceptual models, the introduced conceptual model separates the concept of an agent environment into physical container, social regulation and functional entity, which overcomes the limitations on environmental completeness in other ABM models and provides flexibility in simulating different system scenarios.
The concept of a physical container is introduced to represent the physical environment where agents stay. Typical physical containers are school, classroom, office, bedroom, train, etc. Physical containers are organized hierarchically. Moreover, this concept makes it much easier to include a transportation component in a social simulation model, which is achieved by considering vehicles as movable physical containers in the model.
The concept of social regulation, borrowed from the multi-agent system community, is used to model artifacts that can guide and influence agent behavior globally (rules/norms/institutions). With this concept, agents can respond to different situations during a simulation run. For example, regulating agents’ behavior during a disease outbreak is an indispensable part at a large-scale agent-based epidemic simulation. How agents would respond to interventions during a disease outbreak would have a big impact on the model outcomes.
The concept of functional entity, borrowed fromthe object-oriented paradigm, is used to model the extra objects in the system that can influence or directly change attributes of either agents, physical containers or social regulations. For example, a disease is modeled as a functional entity to change agents’ healthy status. Temperature can be modeled as a functional entity to change the transmission probability of a disease in a specified location (physical container).
Using reference implementations of these concepts, a model of a large-scale artificial city of Beijing is constructed as a case study to test policies for controlling the spread of disease among the full population (19.6 million). This case study can be considered as a proof of concept which exemplifies how large-scale social systems with complex human behavior and social interactions can be modeled with the help of the proposed conceptual model, but still gains reasonable performance. It also indicates potential use in other social science areas, such as microscopic transportation systems and city level evacuation planning.