Its been a while since (to say the least), that we wrote a position paper about agent-based modeling. But with agent-based modeling becoming more widely accepted and the growth of machine learning within the geographical sciences we thought we would revisit some of the existing challenges (e.g. validation, representing behavior) and discuss how machine learning and data might help here. To this end, Alison Heppenstall, Nick Malleson, Ed Manley, Jiaqi Ge and Mike Batty, have recently published a paper entitled "Future Developments in Geographical Agent-Based Models: Challenges and Opportunities" in Geographical Analysis. Below we provide the abstract to the paper, and if this is of interest please follow the links to the paper itself.
Abstract
Despite reaching a point of acceptance as a research tool across the geographical and social sciences, there remain significant methodological challenges for agent-based models. These include recognizing and simulating emergent phenomena, agent representation, construction of behavioral rules, calibration and validation. Whilst advances in individual-level data and computing power have opened up new research avenues, they have also brought with them a new set of challenges. This paper reviews some of the challenges that the field has faced, the opportunities available to advance the state-of-the-art, and the outlook for the field over the next decade. We argue that although agent-based models continue to have enormous promise as a means of developing dynamic spatial simulations, the field needs to fully embrace the potential offered by approaches from machine learning to allow us to fully broaden and deepen our understanding of geographical systems.
Building on past posts about our work with respect to generating large scale synthetic populations for agent-based models, we have a new paper entitled "Integrating Social Networks into Large-scale Urban Simulations for Disaster Responses" that was accepted at the 3rd ACM SIGSPATIAL International Workshop on GeoSpatial Simulation. In the paper we discuss our method to create synthetic populations which incorporates social networks to generate for the New York megacity region. To demonstrate the utility of our approach, we use the generated synthetic population to initialize an agent-based model which not only generates basic patterns of life (e.g., commuting to and from work), but also allows us to explore how people react to disasters and how their social networks are changed by such events.
If sounds of interest to you, below we provide the abstract to the paper, along with our synthetic population workflow and some sample outcomes from the model. At the bottom of the post we provide the full reference and link to the paper (the paper itself also links to a GitHub repository where more information about the synesthetic population can be found).
ABSTRACT:
Social connections between people influence how they behave and
where they go; however, such networks are rarely incorporated
in agent-based models of disaster. To address this, we introduce a
novel synthetic population method which specifically creates social
relationships. This synthetic population is then used to instantiate
a geographically explicit agent-based model for the New York
megacity region which captures pre- and post- disaster behaviors.
We demonstrate not only how social networks can be incorporated
into models of disaster but also how such networks can impact
decision making, opening up a variety of new application areas
where network structures matter in urban settings.
KEYWORDS:
Urban Simulation, Agent-based models, Synthetic Populations, Social
Networks, Geographical Information Systems, Disasters.
Synthetic population and social network generation workflow.
Synthetic population at household level within a census track (A) and social network of one individual (B).
Example of a heat-map of traffic density (A) Manhattan is center of the plot. The impact area of the disaster and the health status of the agents (B).
Full Reference:
Jiang, N., Burger, A., Crooks, A.T. and Kennedy, W.G. (2020), Integrating Social Networks into Large-scale Urban Simulations for Disaster Responses. Geosim ’20: 3rd ACM SIGSPATIAL International Workshop on GeoSpatial Simulation, Seattle, WA. (pdf)
In this study we exploit metadata (and not content) emitted from commonplace workplace
technologies such as calendar and workplace messaging apps collected from a tech
company in order to see how mandatory remote work changed communication patterns and how such data can be used to measure organizational
health. If this is of interest to you, below we provide the abstract to the paper along with some of the results with respect to how meetings and communication patterns changed from business as usual (BAU), pre pandemic to that when people were forced to work from home (WFH). Finally at the bottom of the post we provide the full reference and the link to the paper.
Abstract. During the early months of the COVID-19 pandemic, millions of people
had to work from home. We examine the ways in which COVID-19 affect
organizational communication by analyzing five months of calendar and messaging
metadata from a technology company. We found that: (i) cross-level communication
increased more than that of same-level, (ii) while within-team messaging
increased considerably, meetings stayed the same, (iii) off-hours messaging became
much more frequent, and that this effect was stronger for women; (iv) employees
respond to non-managers faster than managers; finally, (v) the number
of short meetings increased while long meetings decreased. These findings contribute
to theories on organizational communication, remote work, management,
and flexibility stigma. Besides, this study exemplifies a strategy to measure organizational
health using an objective (not self-report based) method. To the best
of our knowledge, this is the first study using workplace communication
metadata to examine the heterogeneous effects of mandatory remote work.
Keywords: Work from Home, Communication, COVID-19, Organization.
Full Reference:
Oz, T. and Crooks, A.T. (2020), Exploring the Impact of Mandatory Remote Work during the COVID-19 Pandemic, 2020 International Conference on Social Computing, Behavioral-Cultural Modeling and Prediction and Behavior Representation in Modeling and Simulation, Washington DC. (pdf)
If you would like a pre-print of paper, just let us know and we can email you one.
While Mesa started off with two students from the CSS program at George Mason University: Jackie Kazil and David Masad it has now grown to include over 70 contributors. In this new paper we discuss the rationale for developing Mesa (see https://github.com/projectmesa/mesa) which arose because there was no framework for easily building agent-based models in Python. Furthermore we discuss Mesa's design goals and its architecture and usage, along with who is using Mesa and extensions to it (e.g. Mesa-Geo, Multi-Level Mesa), finally we conclude the paper with future development directions. Below we provide the abstract to the paper and a selection of figures which highlights Mesa's model components (model, analysis and visualization), how various activation schedules are incorporated within Mesa and an illustration of how these
different schemes impact a model and some examples of Mesa's visualization functionality. At the bottom of the post we have the full reference and a link to the paper.
Abstract.
Mesa is an agent-based modeling framework written in Python. Originally started in 2013, it was created to be the go-to tool in for re-searchers wishing to build agent-based models with Python. Within this paper we present Mesa’s design goals, along with its underlying architecture. This includes its core components: 1) the model (Model, Agent, Schedule, and Space), 2) analysis (Data Collector and Batch Runner) and the visualization (Visualization Server and Visualization Browser Page). We then discuss how agent-based models can be created in Mesa. This is followed by a discussion of applications and extensions by other researchers to demonstrate how Mesa design is decoupled and extensible and thus creating the opportunity for a larger decentralized ecosystem of packages that people can share and reuse for their own needs. Finally, the paper concludes with a summary and discussion of future development areas for Mesa.
Mesa model components: model, analysis and visualization.
Activation schedules within Mesa and an illustration of how these different schemes impact a model. In this case the Prisoner’s Dilemma. Defecting agents are in red and cooperating agents are in blue. Each image is from the same step, but different activation schemes are used.
Model visualization of two Mesa applications within a web browser: (A) Wolf-sheep predation Model. (B) Virus on a network (Source: https://github.com/projectmesa).
Full Reference:
Kazil, J., Masad, D. and Crooks, A.T. (2020), Utilizing Python for Agent-based Modeling: The Mesa Framework, in Thomson, R., Bisgin, H., Dancy, C., Hyder, A. and Hussain, M. (eds), 2020 International Conference on Social Computing, Behavioral-Cultural Modeling & Prediction and Behavior Representation in Modeling and Simulation, Washington DC., pp. 308-317. (pdf)
One thing however, is that while within the ML community at large, it is common to compare different approaches and take the one that gives the best result (e.g., like we did in the Communities, Bots and Vaccinations paper), this is not the case within the social simulation community. There has been little written with respect to why one ML method was chosen over another, or how the simulation results might be different if different ML methods were used. To address this gap we demonstrate the integration of three machine learning methods (i.e., Evolutionary Computing, Q Learning, and State→Action→Reward→State→Action (SARSA)) into the well-known agent based model: Sugarscape (in this instance we modified NetLogo's "Sugarscape 2 Constant Growback"). Our rationale for choosing the Sugarscape
model was that it is well known within the social sciences and as the
purpose of this paper was not to solve or explore a specific social
issue, but to show how different ML methods can be used within the same agent-based model and to
show how different methods impact the results of a model.
If this type of research is of interest, below we provide the abstract to the paper, a flow chart of the model execution along with some results. At the bottom of the past you can find the full reference and a link to the paper. Supplementary material can also be found at https://tinyurl.com/ML-Agents.
At this link, the model presented in this paper along with a full
description of it following the Overview, Design concepts, and Details
(ODD) protocol can be found. We do this to allow others to replicate the
results and adapt the ML methods for their own applications if they so
desire.
Graphical User Interface of the “Creating Intelligent Agents” model. From left to right: input parameters, agents within their artificial world, and aggregate model outputs
Abstract.
Over the last two decades with advances in computational availability and power, we have seen a rapid increase in the development and use of Machine Learning (ML) solutions applied to a wide range of applications, including their use within agent-based models. However, little attention has been given to how different ML methods alter the simulation results. Within this paper, we discuss how ML methods have been utilized within agent-based models and explore how different methods affect the results. We do this by extending the Sugarscape model to include three ML methods (evolutionary computing, and two reinforcement learning algorithms (i.e., Q Learning, and State→Action→Reward→State→Action (SARSA)). We pit these ML methods against each other and the normal functioning of the rule-based method (Rule M) in pairwise combat. Our results demonstrate ML methods can be integrated into agent-based models, that learning does not always mean better results, and that agent attributes considered important to the modeler might not be to the agent. Our paper's contribution to the field of agent-based modeling is not only to show how previous researchers have used ML but also to directly compare and contrast how different ML methods used in the same model impact the simulation outcome, which is rarely discussed thus, helping bring awareness to researchers who are considering using intelligent agents to improve their models.
Mean result for vision for all rule combinations (50 model runs).
Full reference:
Brearcliffe, D.K. and Crooks, A.T. (2020), Creating Intelligent Agents: Combining Agent-Based Modeling with Machine Learning, The 2020 Computational Social Science Society of Americas Conference, Online. (pdf)
