Agent-Based Modeling Suggests We Can Modulate COVID-19 Spread By Encouraging Localized Social Interactions — Part One

In the 5 months since organizing the COVID-19 Simulation Summit, SingularityNET and Singularity Studio have been gradually evolving our software framework for Agent-Based Modeling of the pandemic and leveraging it to run simulations of COVID-19 spread in various contexts. In this post, we describe some of the early results we’ve obtained in this quest, and some of the policy directions they tentatively suggest.

A sequel post gives more in-depth supplementary information, for those who want to dig into some of the simulation results we’ve been looking at here in the SingularityNET research lab.

A more informal take on these ideas and results can be found in my talk in the Medical Applications stream of the D.OS (Decentralized OS Summit) event that SingularityNET and Cardano co-organized on November 9, 2020.

As a number of the talks at the COVID-19 Simulation Summit pointed out, the agent-based simulation paradigm allows a finer-grained sort of modelling than is displayed in the simplistic epidemiological models of COVID-19 spreading dynamics that are driving current policy decisions. Most critically, the commonly used simplistic models fail to account for the different behaviour patterns of different classes of people — that is, they leave out the social element, which is accounted for directly in the agent-based approach in which a region (or the world as a whole) is modelled as a specific geometry occupied by interacting autonomous agents with a diversity of specific behaviour patterns.

Our SingularityNET Agent-Based Modeling (ABM) framework can be used to construct multiple types of simulation models, including fine-grained models of particular cities, states or nations, or more universal models of generic simulated environments intended to highlight general properties of COVID-19 dynamics in various sorts of real-world situations.

The simulation results we’ll describe here are focused on demonstrating the subtle dependencies of COVID-19 spread on the social interaction patterns of the people in a particular region. Among other things we will present simulations demonstrating how the conditions for avoiding hospital overload, and for achieving herd immunity to COVID-19, depend on cultural patterns of social interaction — and in particular on how tribalized or localized peoples’ social interactions tend to be.

What we find is that, separately from public health habits like mask-wearing, social patterns like the typical “cliquishness” of peoples’ social interactions can make a huge difference in epidemiological parameters like the amount of the population that needs to be immune to yield herd immunity, and the maximum number of people infected at any given time, which is critical to control for avoiding hospital overload.

These results highlight, among other factors, the generally underappreciated subtlety of the phenomenon of “herd immunity.” Most epidemiological analyses of herd immunity are based on simplified assumptions such as a population with homogeneous susceptibility to infection, and uniform social mixing within the population. When one replaces these assumptions with more realistic ones, one discovers that herd immunity is a much more dynamic and pliable phenomenon, which is influenced by physiological heterogeneity in the population, by patterns in social interaction dynamics, and by policies like lockdowns including sometimes the specific dates with which lockdowns being as well as their durations.

A lockdown, among many other factors, can actively shift the dynamic equilibrium which is herd immunity — herd immunity is not just something obeying quantitative rules that apply independently of policy decisions such as lockdowns. You are right. If you look to the right of these charts, no more persons get infected, even though there are more available to be infected, and that point depends on the connectivity that we just changed with the lockdown.

We are certainly not the first to explore the subtle dependencies of herd immunity on diverse factors — for instance, an excellent article from Atlantic Monthly earlier this year explored nonlinear-dynamics epidemiological models showing that if peoples’ susceptibility to infection is heterogeneous then the herd immunity threshold can be as low as 20%, While most epidemiologists — as well as journalists, politicians and so forth — have been thinking about herd immunity in a simplistically fixed way, our work is in the tradition of the smaller community of chaos-theory-savvy epidemiologists who see herd immunity from a richer perspective.

Furthermore, the dependencies we find in our simulations between social properties like clumpiness and epidemiological measures like maximum infected population and herd immunity threshold are not linear. Rather, our simulations reveal “threshold” type behaviour, wherein once clumpiness gets above a certain critical level, then COVID-19 spread control suddenly becomes significantly easier. The existence of this sort of threshold in real-world systems has yet to be validated but seems perfectly plausible as many complex, self-organizing social-network phenomena display similar properties.

Our simulations suggest that, if clumpiness of social interactions is above a certain threshold, then it may be perfectly fine for policy-makers to be a little laxer regarding triggering and duration of lockdowns. (And based on our limited but highly suggestive work so far, this threshold doesn’t look ridiculously high, it seems we are operating with a pragmatically feasible regime here.) With below-threshold clumpiness, the impact of lockdowns seems to depend more acutely on how sensitively they’re triggered and how long they last. With above-threshold clumpiness, it seems not to matter quite so much exactly how many cases there need to be to trigger a lockdown, nor exactly how long the lockdown lasts.

There is a simple practical message here for COVID-19 public health policy-makers: Make policies that, as much as possible, guide peoples’ social interactions into largely-discrete “clumps” or subgroups. For COVID-19 pandemic management, tribalism is your friend!

A simple way (though not the only way) to achieve the desired clumpiness would be to guide peoples’ social interactions toward physical localization — if people are mainly interacting with others from their local geographic area, then this is one way of achieving social-interaction clumping.

For instance, opening elementary schools which mainly serve students living in the same local neighbourhood is probably much less problematic in terms of overall COVID-19 spread than opening high schools or colleges that aggregate students from a wider region.

And if people could be incented to eat out at restaurants near their homes, COVID-19 spread could be significantly attenuated relative to situations where they are travelling further to eat out (and thus mixing with a broader variety of people). The same would hold for in-store shopping. One can easily imagine non-oppressive ways to incent this sort of behaviour — e.g. give people Public Health Reward Points for dining or shopping near their homes (and of course given even greater reward points for ordering take-out food or getting home delivery of groceries), where Reward Points can be used for discounts or upgrades similar to frequent flier miles.

A bit of creativity in designing and implementing these sorts of policies seems warranted, given the incredible economic and human cost of the current crude and simplistic policy approaches, which also are not working terribly well in the US nor in many other Western countries.

Along with their potential immediate practical interest, these preliminary simulation results also illustrate the importance of considering all the relevant factors — including “soft” factors like cultural patterns of socialization — in doing the modelling used to drive policy decisions.

The underlying philosophy driving this work is: If we’re going to get through this pandemic while minimizing loss of life and also minimizing collateral economic, social and psychological damage, we will need to be clever and artful at choosing policies that match the particulars of the situations in particular regions. As these early results show, Agent-Based modelling can be a powerful tool for executing this sort of artistry.

Overall, if our pandemic management policies were based on agent-based simulations of COVID-19 spreading, incorporating the different economical, geographic and social characteristics of different regions, then we would almost surely be experiencing significantly lower rates of infection and death than is currently the case in most Western countries, along with much less disruption to practical life.

A Simple(-ish), Generic COVID-19 Simulation Scenario

Having given you the TL;DR summary of some of the key practical take-aways from our latest batch of COVID-19 simulation modelling results, let me take a step back and explain a bit more about how the sort of modelling we’re doing works.

What goes into an Agent-Based Model of COVID-19 spread? Even a relatively simple and generic model that doesn’t aim to simulate a particular real-world region, but rather to explore more general properties of infection spread, still need to include quite a lot of factors in order to be realistic enough to be meaningful.

Here are some of the key factors we’ve incorporated into our models so far:

Multiple classes of individuals:

• Infants and toddlers
• K-12 school-aged children
• Working-age adults
• Elders

A variety of sub-classes for working-age adults:

• Office workers (have a work location but can work from home)
• House-bound people (whether they work from home, are trophy househusbands, etc seems not to matter much)
• Factory workers (can’t work from home but don’t meet non-coworkers during their work hours)
• Retail workers (can’t work from home AND meet non-coworkers during their work hours)
• Healthcare workers and other essential workers
• Transportation workers (no fixed work location)

Economic impact modelling that factors in the above classes and sub-classes. (A future to-do is to extend this economic model to factor impact on businesses, and estimate layoffs and closures from that.)

Classes of locations such as:

• Houses
• Apartments in a building
• Offices in a building
• Shops
• Schools
• Hospitals (here it becomes key to model the capacity of the hospital system: regular beds and ICU/ventilator beds)
• Factories
• Fun gathering spots

Individual health attributes — most simply modelled via an overall health parameter, which can be reduced by the presence of comorbidities and immuno-suppressed individuals and increased for fit individuals.

We have also incorporated into our models a reasonable diversity of personality characteristics on the part of the simulated agents. For instance, assuming everyone will follow social distancing guidelines isn’t realistic, but randomly deciding who doesn’t is also bad. To obtain the individual behaviour we expect, each simulated individual can be endowed with properties like:

• A risk tolerance parameter
• A “herding behaviour” parameter that makes an agent more or less likely to adopt a given behaviour (in compliance or defiance of official policies) based on the behaviour of their peers.

We have also considered explicitly modelling polarization, where we give agents affiliation groups, and, agents imitate affiliated ones and distance behaviour from ones with dissimilar affiliations — but this has not yet been worked into our practical simulations.

All this forms the framework upon which one can implement a variety of social policies: complete lockdowns, age-based lockdowns, partial lockdowns, social distancing mandates, etc.

At the moment the agents in our simulations are not learning and reasoning, but we have architectured our simulation framework with the introduction of more intelligent and adaptive agents in mind; and when the time comes, these personality parameters will form key inputs to each learning agent’s utility function.

Modeling Social Interaction Clumpiness

To abstract away from the particular geography of any specific town, city, state or nation, we have done many of our initial simulations using a simplified geometry: A town consisting of a square grid of “home districts”, overlaid with school and work districts. By controlling how many home districts are covered by a school district or a working district, we can control aspects of the social interaction pattern of the individual agents living in the homes and travelling to the schools or workplaces.

The more home districts are covered by a school district or work district, the less “clumpy” is the typical social interactions of the agents — i.e. the more far-flung are the homes of the other agents that a given agent interacts with. If the school and work districts are small, then each agent will be mostly interacting with a fairly small group of others who live nearby — very “clumpy” interactions.

In this simple model, the clumpiness of social interactions may then be experimented with along with numerous other parameters describing the agents and their behaviour — such as e.g. their age distribution, their propensity to wear masks, and so forth.

This is what led us to findings such as the ones I summarized above — that the impact of lockdowns with various different triggers and lengths depends sensitively on the degree of “clumpiness” of social interactions in a population — where clumpiness is defined as the degree to which people interact socially within specific social groups they belong to, versus with broader members of the public.

For instance, in our simulations: In a population where interactions are sufficiently clumpy, then one can be laxer regarding when lockdowns are triggered and how long they last, and still avoid overwhelming the medical system and keep the herd immunity threshold relatively low. But in a population where interactions are not clumpy enough, one may need to be quite sensitive regarding triggering lockdowns and keeping them in place for a while.

This example illustrates the subtle dependence of critical aspects of COVID-19 spread on contextual phenomena such as the cultural socialization habits of the people in a certain region.

The sequel blog post to this one takes an in-depth look at some of the simulation results behind these conclusions regarding social clumpiness and lockdowns — i.e. lots of graphs and charts from various simulation runs!

The policy implications of this particular finding are not hard to see, and were hinted at above: Anything that can be done to incent people to direct most of their social interactions within limited-size “clumps” will nudge the spreading dynamics in a more tractable direction. Increasing social-interaction clumpiness will make it easier to control spread using lockdowns — making it easier to keep maximum infection rate at a non-catastrophic level in terms of maximum hospital capacity, and making it easier to achieve herd immunity with less widespread immunity.

Toward Finer-Grained Simulations

The Agent-Based Models we’re currently working with occupying a sort of middle ground between the extremely simplified equational models that most epidemiologists are working with, and fully detailed models of particular towns, cities, countries, etc. All of these various modelling approaches at their different degrees of granularity have their own value and importance.

ABM models at the level of granularity we’re exploring now are well suited to understanding the fundamental nonlinear dynamics via which infection spread interacts with broader social and psychological factors — something that we believe is critical to account for in policymaking.

Finer-grained models of specific jurisdictions and regions will also be extremely valuable and will give much more precise policy guidance, and these will most naturally be constructed as extensions and elaborations of the more generic ABMs we’ve been exploring so far. One expects that finer-grained and more situation-specific ABMs will display the same general dynamic phenomena we see in our current simulations — but with quantitative differences that may sometimes be quite impactful on the policy level.

The underlying simulation modelling framework we’ve built and used for the simulations reported here can be much more broadly, including for finer-grained simulations. It can be overlaid on maps of specific cities, states or nations and then used to simulate detailed COVID-19 dynamics within these specific regions, based on detailed local data.

We Should be Guiding COVID-19 Policy Using Realistic Agent-Based Models

Given that we in the West have mostly avoided the hard lockdowns and thoroughgoing contract tracing approach that worked so well for combating COVID-19 in many Asian countries, in favour of a more locally varying and heterogenous approach — it would be great if we could shift to a more intelligently locally varying and heterogeneous approach.

The cost of configuring and running agent-based simulations of various different jurisdictions and regions, while not zero, would be surely far smaller than the human and financial cost being incurred right now by the current melange of crude policies whose details are being driven largely by unscientific intuition and political bargaining.

The basic public health methods being pursued right now across the West are not wrong, in terms of the goal of keeping COVID-19 spread down to a level where the available medical infrastructure can humanely deal with it: advise or mandate mask-wearing in shared indoor spaces, encourage social distancing, close down public social interactions when it’s possible to do so without excessive human or economic cost. But there are a lot of devils in the details, and politics and common sense are not sufficient tools for wrestling with these devils. Agent-based modelling is not an all-powerful tool either, especially given the various limitations of available data, but it can be a very valuable supplement to other methods and we believe it should be far more widely used.

Join Us

We hope you will join us in our mission to democratize AI and to use its transformative powers to enact real, positive and lasting change. The best way forward is to come together and work practically towards creating a better future.

If you are looking to monetize your AI services or create new ones, we invite you to learn more about the nature of our platform and what it has to offer by visiting the SingularityNET developer portal.

We are proud of our developers and researchers that are actively publishing their research for the benefit of the community; you can read the research here. For any additional information, please refer to our new roadmap. To stay informed about all of our developments, please subscribe to our newsletter. You can connect with our passionate and talented community by visiting our community forum.