Novel Concepts for Agent-Based Population Modelling and Simulation: Updates from GEPOC ABM

TL;DR

This work advances large-scale agent-based population modelling by introducing a birthday-centered time-update, a co-simulation-inspired simulation-layer for parallelisable time-continuous updates, and a data-driven parametrisation workflow anchored in open census data and life-table theory. The methodology couples a two-layer conceptual model with Farr-based probability calculations and robust resolution-disaggregation techniques to produce accurate, open-data–driven population projections, demonstrated through a Austria-wide validation against census and forecast data. Results show high fidelity over 2000–2025 and informative behavior against 2025–2050 forecasts, while identifying limitations related to data resolution and disaggregation. The innovations are presented as transferable to other large-scale ABMs and applicable across epidemiology, climate-disaster analysis, and related decision-support contexts, offering a practical pathway to scalable, transparent population simulations.

Abstract

In recent years, dynamic agent-based population models, which model every inhabitant of a country as a statistically representative agent, have been gaining in popularity for decision support. This is mainly due to their high degree of flexibility with respect to their area of application. GEPOC ABM is one of these models. Developed in 2015, it is now a well-established decision support tool and has been successfully applied for a wide range of population-level research questions ranging from health-care to logistics. At least in part, this success is attributable to continuous improvement and development of new methods. While some of these are very application- or implementation-specific, others can be well transferred to other population models. The focus of the present work lies on the presentation of three selected transferable innovations. We illustrate an innovative time-update concept for the individual agents, a co-simulation-inspired simulation strategy, and a strategy for accurate model parametrisation. We describe these methods in a reproducible manner, explain their advantages and provide ideas on how they can be transferred to other population models.

Figures (6)

• Figure 1: Visualisation of the birthday-problem. Depending on when in the course of the year an event such as birth, migration or death occurs, it will be counted towards different entries in the census data for that year.
• Figure 2: Overall update scheme of GEPOC ABM. The time advancement is sketched with solid horizontal arrows, event-scheduling with dotted harpoons. The simulation-layer acts as runtime-infrastructure for the discrete-event simulations (DESs) of the individual agents. In every macro step, the simulation-layer (blue) first observes the individual states of the DESs (upwards arrows) and potentially interferes with them (downwards arrows) which includes adding or removing events (blue dotted arrows) or entire agents. The red dotted arrow displays an interaction between agents via the simulation-layer -- in this case, a "birth"-event.
• Figure 3: Event-graph representation of the birthday-centred update-scheme for agents. A Birthday event regularly schedules itself every year. On its occurrence, demographic events (e.g. Event) taking place in the course of the agent's next life-year will be planned. They will be scheduled with a certain probability at a random point within the next life-year.
• Figure 4: General parametrisation and validation scheme of GEPOC ABM. Population, death, birth and migration counts in form of census counts (white) and forecasts (grey) pose the source for the parameter values and the validation reference at the same time.
• Figure 5: Simulated census (full lines) compared to the reference data (dashed) for different demographic indicators between 2000 and 2050. Panels [a] and [c] show the comparison of the population counts, differentiated by male and female, [b] and [d] show total births, deaths and emigrants. While panels [a] and [b] display total numbers, panels [c] and [d] show the differences together with the maximum and minimum deviations $e_{max}, e_{min}$. The grey background indicates years in which the reference series consists of official forecasts; the white background indicates years with census counts.

Theorems & Definitions (1)

• Definition 2.1: Probability of Death