Agent-Based Modeling for Pandemics: The COVID-19 Simulation Summit, Part One

Original article was published on Artificial Intelligence on Medium

Agent-Based Modeling for Pandemics: The COVID-19 Simulation Summit, Part One

Covid-19 has changed the face of the world over the last half-year, and as we move into the second half of 2020 the year looks very different than we thought it would in 2019. Even though the disruption this pandemic has caused touches every area of our lives, it is, however, encouraging to think that humanities ability to adapt has ensured our survival throughout history. In fact, the idea that major developments in our evolution happened not gradually, but in fits and spurts during periods of global upheaval is not without merit.

On April the 30th 2020 in partnership with DAIA — The Decentralised AI Alliance, we ran an online virtual summit centred on how AI and agent-based modelling can help during pandemic situations, featuring world-leading experts in the fields of healthcare, epidemiology, Artificial Intelligence, blockchain and statistical analysis In this two-part blog post, we will summarise the talks delivered.

The Simulation Summit Opening and Keynote — Dr Ben Goertzel

Agent-based modelling is an interesting way of simulating complex biological and social situations, and detailed computer simulation models of aspects of society exist today in various forms — from computer games like SIM City to serious in-depth applications in epidemiology, military intelligence, government, and economics.

As an approach, agent-based modelling is preferable to setting up high-level differential equations to describe a system, because greater intricacy of behaviour is observable by programming software agents with rough incentives to interact with each other. Coupling agent-based modelling with new technologies like blockchain is especially exciting, for example, blockchain can help with collecting data to feed into simulation models in a way that respects privacy and data sovereignty.

Much of this issue regarding COVID-19 is that social policy decisions regarding the control of COVID-19 are currently being made based on very simple modelling and analysis methods. With trillions or tens of trillions of dollars, at stake policymakers are making substantial decisions with huge impacts based on limited information. Data is currently available that will allow for a more fine-grained approach to modelling the spread of COVID-19, and it is in all our interest to pursue more accurate and in-depth modelling of the disease’s spread.

Agent-Based Modelling can potentially help us answer the following types of questions:

  • How long does social distancing need to stay in place in a certain region to really have its impact?
  • If you let a certain class of people social distance less than others say, schoolchildren, how much impact is that going to have?
  • How much impact would you have on reducing the death toll by more effectively managing nursing homes and more effectively keeping the elderly and immunocompromised?
  • How useful are our travel bans really when the travel is between places that already have a lot of COVID-19?
  • When do you open school? Do you? Do you allow children to go to school only for 4 1/2 days?

No one has done a careful modelling analysis of the costs and benefits of these different approaches yet, and while agent-based modelling cannot do everything — it can still answer a lot of questions.

Modelling how the disease will spread based on gradual re-opening of our society can help policymakers understand what effects their decisions could make on their complex societies. While simpler methods are more commonly used for modelling, when so much is at stake there is no reason not to take a more fine-grained accurate approach.

It is also important to note the same agent-based modelling framework can be used for future pandemics, and other situations besides pandemics, because an in-depth agent-based model of society has a variety of applications. Building agent-based models of COVID-19 could lay the foundation for future studies/development of more intelligent agent-based modelling, that is suited for a variety of purposes.

The Coronavirus and Modeling of its Pandemic — Dr Yaneer Bar-Yam

The policy has real-world implications and we have to think about them carefully. Keeping this in mind, how do we construct effective models?

Complex systems science is about dealing with problems that don’t fit in simple mathematical paradigms, like calculus and statistics. The reason we need to shift away from these things is that behaviours that exhibit sharp transitions cannot be modelled very well with formal math. It is not that the math is incapable of modelling behaviour, it is that the variables we often choose are not well selected and the past data we are looking at is irrelevant to the situation at hand. The real imperative behind complex systems is figuring out what the right variables are and which are the most important variables in the problem.

Regarding COVID-19 — fatality rates, speed of spread, geographic density are all important variables that need to be considered by policymakers, and agent-based simulations are potentially helpful here because they can show policymakers how they impact COVID-19’s exponential growth. We know how to stop outbreaks — essentially you must stop transmission, and if successful the multiplicative number of cases going up becomes a multiplicative going down. This is the game plan for getting rid of disease and has successfully been used to stop multiple Ebola breaks. Understanding this policy answer should be obvious, only once the disease is stopped can normal activities be resumed. However, there is always a huge amount of confusion due to people speculating about things and we often lose clarity about what is really important.

When studying host-pathogen systems, which are very simple models that consist of prey and predators, it is imperative that you have a good spatial model to model phase transitions. For instance, having a good long-range transportation spatial model is important as systems transition from having local outbreaks to having a global pandemic. Statistics are not helpful when phase transitions move towards dramatic changes that have not been seen within datasets before, and statistical models cannot tell you the effects of an entire societal behaviour change.

While contact tracing is essential, how are you supposed to do it efficiently in Monrovia, in Liberia, or a major city like New York? The answer turns out in complex systems and scale — you cannot address the problem at the individual level anymore, but have to address it at the community level. A community-based complex systems analysis that acknowledges community-wide behaviours and the effects of them is needed. Behaviours such as social distancing, making sure people wear masks, and other actions can be modelled into complex agent-based simulations and then be used to inform policymakers.

Current approaches that use calculus and statistics cannot inform policymakers what will happen if a new and novel behaviour is widely implemented in society. However, an agent-based simulation approach can show the policymaker a high-level overview of their community — and then what happens to the community tomorrow, one week from now, one month from now if a certain behavioural change is implemented broadly throughout their community. This type of understanding of the “effects” of a policy is missing in a purely statistical or mathematical approach and is a unique property that makes agent-based simulations an essential part of a policy maker’s toolset.

Variolation and Vouching as Alternate Pandemic Policies — Dr Robin Hansons

The usual right thing to do with a pandemic is to catch it early and lock it down strong, and the usual result from this approach has been a success. We seem to have failed, at least initially, at doing it the right way — we did not catch COVID-19 early and we have not locked it down very well. And even though there have been these models from other nations who have seemed to have done it successfully, we refuse to just take their whole package of policies that worked and copy them wholesale. Instead, we insist on remixing our own invented new solutions based on our political constraints and tastes and everything else. Especially here in the United States, we are so far failing.

People talk about flattening the curve, but we should perhaps also be working on a plan B where we plan for the scenario where we fail to contain COVID-19. This is not a first-best solution but it may be necessary to consider and ask how to best handle such a plan B. Often people talk about flattening the curve and there is a great benefit from preventing the medical system from being overwhelmed all at once. However, what if the deliberate infection was considered a possibility. For instance, imagine there was a hotel where you pay to go to and the hotel tries to infect you as soon as you get in. Once infected — you have to stay in the hotel until you seem to be recovered. This way you can choose exactly who is infected and so you can use isolation resources very efficiently.

Deliberate infection may give a factor of a third or 50 per cent reduction in death rates due to controlling who gets infected, but it turns out there is actually a much bigger reason to consider deliberate infection. With deliberate infection, you can actually control the dose, the vector of delivery, and perhaps the strain — and through these through controls, you can further dramatically lower the death rate. For instance, systems have shown that if you are initially given a small dose of a virus, your immune system has the potential to have a head start building up immunity against it. Further, since COVID-19 affects the lungs significantly — it may make sense to deliberately infect through the skin or intestines to improve chances of recovery. Thus, deliberate infection with controlled dosage and delivery could build up a herd immunity safely and quickly. What is needed is an initial trial, of maybe 100 people, but the problem has been getting medical ethics permission due to infection with a live virus.

However, if we are facing the prospect of most of us getting infected anyway, then we should be prepared with a plan B — should it look like we are no longer able to contain the virus. The key problem, as economists see it with a pandemic is the externality of infection. By infecting someone else, you affect them, you make them worse off, and you don’t have a sufficient incentive to pay attention to that. So when everybody knows there’s a pandemic going on, people have a sufficient incentive to isolate themselves, to prevent themselves from being infected by other people, but they do not have a sufficient incentive to additionally try to prevent other people from being infected by them.

We often have a choice in dealing with externalities, one of them is regulation and another is a liability. So the question is, could we use liability instead of regulation for pandemic infections? The potential gain from regulation is that it is a one-size-fits-all approach that is easy to implement, albeit crude. If we want a very context differential dependent approach — we should use liability. Liability gives the incentive to pay very close attention to context when making decisions in the pandemic — should you go to work, should you wear a mask, have you had symptoms lately, should you social distance? Our crude regulations are not making these distinctions and we are lacking fine-grained incentives for people to pay close attention to the chance that they might infect someone.

As with auto insurance for liability — infection liability insurance could also become a requirement. If you want to leave your house you need some infection liability insurance and some insurance company guarantees that if you infect someone they will pay. The insurance company may be stuck with some conditions, e.g. make sure you wear a mask and socially distance, require GPS tracking, install an app on your phone, etc. Infection insurance acting as a passport for being immune seems like a more general robust solution to these sorts of pandemics, which allows a lot of local behaviour to adapt and does not require that you wait for the government to figure out the right thing to do. In conclusion, there are many approaches to a Plan B and we only invoke plan B when plan A seems to fail, but we need to rapidly have a plan B ready.

Agent-Based Simulation of COVID-19 Health and Economical Effects — Dr Petronio Candido

There are a number of people forecasting the number of COVID-19 infected and its fatality rate, and it is clear that the agent-based framework proposed will allow for dynamic, and robust optimization and decision making.

We know the primary causes COVID-19, but what are the secondary causes? For example, the environmental figures such as temperature and the humidity are affecting the rate of contagion and magnified by political misconceptions and wrong decisions of the rulers. Some social habits and misconducts. To model secondary and primary causes of COVID-19 we want to make a model based on data sets and on the knowledge base we have about the disease. Once we build creative models, we can test interventions that will be applied to the pandemic. These models are functions that have steps to generate new states, which are continually changing. It is important to define a parameter set that is composed by a set of constants and a set of decision variables that can be changed. The main loop of the simulation is the continuous updating of this internal state given the set of the parameters and the response variables.

The general goal of such a model is to help the government how to help the whole government prescribing interventions. The global state of the system is an aggregation of individual agents, and each agent in the simulation needs to be given individual parameters and actions in their individual state. Furthermore, there is a common environment that has shared parameters which also interacts with agents. Inside this environment is where agents interact with each other, come into contact, and perhaps get infected. In the first iteration of our simulation, we have a set of data, a set of parameters as input, and agents. Agents change their location within their shared environment and the simulation tries to account for the movement of agents contacting each other and the spread of the contagion. The agent-based simulation is a population of agents living inside an environment in a continuous loop — by initializing the shared environment, given parameters, we are able to initialize the simulation of a potential population.

Our goal is to predict the impact of the social interventions in the epidemic and slowly make it as holistic as possible. We have an environment space — that elaborates on population size and density — and we select the number of iterations the simulation should run for and its time horizon. Then it is important to input scientific knowledge into the simulation — for instance, it is established that the severity of COVID-19 changes according to age and these items of information need to also be included in the simulation. With all this information coded in — we can then see what happens as we decrease mobility, put people in quarantine, or let people freely move about. Additionally, we can test the capacity of the health system and the effects of expanding health facilities.

Agents have a degree of freedom in their movement and we need to model these variable movements. Agent movement is necessary for the circulation of money, goods, services, which unfortunately also makes the virus circulate. If the agent is alive we assume that it will randomly move within the environment, model their interactions with other agents, and then check the internal state of the agent after its movement (what economic/health changes have happened to the agent through the course of their movement). From this information, we can model where the number of hospitalization cases crosses a critical line, while also modelling scenarios where we restrict mobility and their effects on other spheres like the economy. More sophisticated things can be tested such as two populations with different policies interacting with each other, vertical isolation (the isolation of specific age groups), and how the pandemic will progress if contagion/fatality rates change.

Modelling and Evaluating Intervention Options and Strategies for COVID-19 — Dr Eva Lee

Agent-based simulations can input information from a variety of sources such as census data and citizen on the ground data. Agents can scale from the micro to macro level, and in a simulation, you can control a single agent, an individual, a family, a group of people, a business, small business, or big business at different levels of granularity. The challenge is that simulation is really computationally intensive, but it can help with optimizing scenarios on the ground when there are scarce resources.

Numbers on the ground are the best estimates we have to model what is going on, and we need these figures to determine COVID-19’s contagious spread. There are numerous challenges with COVID-19, for instance, we know that there are asymptomatic and symptomatic patients, different incubation periods for those infected, affects various demographics differently, and can potentially undergo significant mutation.

To successfully deal with COVID-19, we first need the ability to rapidly test for the disease. After testing is established, the next step is to divide and conquer — we want to decentralize the decision making on school closures and mobility restriction to small pockets, meaning that if there is an infection it is small enough for a community to handle it. However, we currently have a testing problem and we are unable to overcome this.

The solution may be strategic testing in an effort to contain the disease, where we rely on hospitals’ ability to separate patients with different coding and staging to direct testing and use remote solutions like telehealth to target areas to test. By identifying health personnel and utilization, which hospitals have a poor understanding of, a simulation can generate an idea of the disease’ progression. If utilization is tracked and hospitals can see daily increases/decreases in hospital inpatient areas you can begin to model it. Societies can use these simulations to optimize resources and minimize total infection and mortality rates.

Agent-based modelling cities and jurisdictions is important because that is where the spread happens and that is where data matters a great deal. Modelling Hong Kong spread is substantially easier than modelling spread in the United States (which is more similar to modelling the spread in San Diego) and you also have better data at the city/jurisdiction level. Open-source data, such as labour information, can be pulled and matched up to specific jurisdictions and used by simulations to model disease spread. This will give policymakers within these jurisdictions the ability to see contingencies at a high level and be able to make decisions based on digestible information.