Organisers: Robin C. Purshouse (University of Sheffield), Ivan Garibay (University of Central Florida) and Joshua M. Epstein (New York University and Santa Fe Institute)

Overview

The generative, or mechanism-based, approach to modeling of social systems uses agent-based models (ABMs) to ‘grow’ the phenomenon under investigation. The modeler designs and implements the ABM, and chooses its parameters and initial conditions (i.e., inputs). Then the model is run to generate an emergent output – if this output can, in some sense, reproduce the phenomenon then it becomes a candidate explanatory model; otherwise it is rejected.

Whilst the ABM community is now focusing heavily on methodology for the consideration of model inputs (e.g, calibration techniques), surprisingly little attention is given to the consideration of model structure – i.e., the nature of the entities and equations in the ABM.

Whilst initiatives such as the Overview, Design concepts and Details (ODD) Protocol encourage modelers to articulate ABM structure in a thorough manner, these initiatives do not stimulate scientific consideration of the plurality at the heart of model structure selection decisions.

Where does a particular structure come from? Does it arise from the art of the modeler, or does it arise from a scientific process? How does ABM speak to theory, and vice versa? How do we choose between alternative mathematical and computational realizations of a specified mechanism, and when do we know that a mechanism can be rejected? What is special about the structure that has been identified, compared to the universe of other structures that could have been chosen? Do multiple, meaningful candidate structures exist and, if so, do these share any similarities

 To embark on the journey to answering these questions, the ABM community now needs to bring the issue of model plurality, and methods for model discovery, to the forefront.

This workshop will introduce the philosophy and ideas that underpin the concept of model discovery, most recently articulated by Epstein as ‘inverse generative social science’, but which have also arisen under the designations of ‘model crunching’ and ‘structural calibration’; related ideas have also arisen in the areas of pattern oriented modeling, model alignment, and model replication and breaking.

The workshop will then set out recent developments in computational intelligence and machine learning methods that have been harnessed for computer-aided model discovery, including how to conduct an efficient search over the wide range of ABM entities and equations that might explain a target phenomenon, and how to use the results to perform abductive inference of key causal mechanisms.

The workshop will then introduce recently developed tools that enable these methods to be used by the ABM community – including object-oriented software architectures and synthetic agent populations, NetLogo and Repast implementations, and integrated software platforms for model discovery and inference. Next, the workshop will present learnings from a diverse set of recent applications of model discovery, including civilization growth and decline, alcohol use in US society, and message cascading on social media platforms.

Finally, the workshop will set the stage for a discussion on the potential of model discovery as a new grand challenge for the ABM community.