Epidemic Modeling - Networks, Mutations, and Waves

Learning / Complexity Dynamics And Control / Agent-Based Modeling / 010

LESSON

Agent-Based Modeling

010 30 min intermediate

Day 346: Epidemic Modeling - Networks, Mutations, and Waves

The core idea: epidemic models are state-transition systems over contact networks, so waves and variants make sense only when you track who can infect whom, how fast state changes, and how the contact graph keeps shifting.

Today's "Aha!" Moment

In the previous lesson, Harbor City treated beliefs as the thing that moved through the fleet. Today the city is facing a propagation problem that is less negotiable. A norovirus cluster has appeared on the commuter ferries that connect the marsh, the container pier, and the downtown terminal. The operations team wants one answer: do they cut routes, rotate crews, or keep service running and intensify cleaning? The answer depends on mechanism, not on a headline case count.

If you only plot total infections per day, the outbreak looks manageable for the first few days. But the downtown terminal is a hub. Crews switch there, passengers transfer there, and cleaning staff touch multiple vessels in one shift. Ten infectious people scattered evenly across Harbor City are not the same as ten infectious people concentrated around that hub. Epidemic modeling starts when you stop treating "cases" as a citywide blob and start modeling how state moves through a network.

That is also why mutations and waves matter. A slightly faster variant introduced on a low-traffic route may die out. The same variant introduced on the crew-exchange loop can replace the old strain in a week. A wave is not just "more cases later." It is what you see when transmission, immunity, behavior, and network structure keep feeding back into one another. The useful mental shift is that epidemic curves are outputs. The model lives underneath, in contact opportunities and transition rules.

Why This Matters

Production teams use epidemic models when they have to act before the system becomes obvious. Harbor City's transit authority cannot wait for hospital admissions to confirm what the ferry network is already doing. It has to decide how many reserve crews to hold back, whether to suspend one transfer route instead of all routes, where to place testing capacity, and when cleaning intervals are worth the lost service time.

Without a model, organizations often overreact to lagging totals or underreact to dangerous structure. A flat citywide average can hide one high-degree terminal that is seeding the rest of the network. A falling case count can look reassuring while a new variant with shorter generation time is already taking over. A useful epidemic model does not predict the future with certainty. It makes the causal levers visible enough that operations, public health, and logistics teams can compare interventions before paying for them in the real world.

This matters outside public health too. Malware spread, crop disease, and supply-chain contamination all involve the same question: what local contact pattern produces the global wave? Epidemic modeling is one of the cleanest places to learn that systems lesson because the state transitions are concrete and the consequences of getting the mechanism wrong are immediate.

Learning Objectives

By the end of this session, you will be able to:

  1. Explain why compartment models are useful but incomplete - Describe how susceptible, exposed, infectious, and recovered states turn raw case counts into a causal model.
  2. Analyze how network structure changes transmission - Identify why hubs, repeated contacts, and route topology create superspreading and uneven risk.
  3. Evaluate how mutations and behavior produce waves - Compare interventions when transmissibility, immunity, and contact patterns are all changing at once.

Core Concepts Explained

Concept 1: Epidemics are state machines before they are charts

Harbor City's first mistake would be to treat every infected rider as identical and every day as interchangeable. In reality, people move through disease states. A rider can be susceptible in the morning, exposed after a crowded transfer, infectious two days later, and recovered the following week. A simple SEIR model captures that progression with four buckets: susceptible (S), exposed (E), infectious (I), and recovered (R).

That structure matters because each bucket changes what the system can do next. Susceptible people can be infected. Exposed people have already been infected but are not yet spreading. Infectious people generate the next round of cases. Recovered people temporarily stop participating in the transmission chain. Once you model those states explicitly, the outbreak stops being a mysterious curve and becomes a transition system with rates you can reason about.

For Harbor City, a coarse daily update might look like this:

new_exposed = transmission_rate * contacts(S, I)
new_infectious = exposed_people / latent_period
new_recovered = infectious_people / infectious_period

The power of this abstraction is speed and clarity. If the latent period is long, today's exposure spike will not show up as today's infectious spike. If the infectious period shortens because infected crews isolate earlier, the same number of exposures can produce a smaller secondary wave. The trade-off is compression. A compartment model is fast enough to run many scenarios, but by itself it does not know that ferry cleaners meet many routes while passengers usually touch only one.

Concept 2: The contact network determines where the outbreak accelerates

Now add Harbor City's route graph. The downtown terminal connects six ferry lines. Night crews rotate through the maintenance dock before morning service. School commuters pack the east inlet line for forty minutes, while the industrial pier line has fewer riders but a stable crew that works together for entire shifts. Those details are not decoration. They define the contact network over which infection moves.

East Inlet ----\
Marsh Point ---- Downtown Terminal ---- Container Pier
Airport Dock ---/           \
                           Maintenance Yard

In a network model, an infectious person does not contribute to transmission everywhere equally. They contribute where edges exist and where contact intensity is high. A deckhand who works the downtown terminal and the maintenance yard is more important epidemiologically than three riders who each take one isolated trip. Superspreading often comes from structure as much as from biology: central nodes, repeated high-duration contact, and bridging roles that connect otherwise separate groups.

This is where R0 becomes operationally useful only after refinement. A citywide average reproduction number can say the outbreak is barely growing, while the effective reproduction number around one hub is well above 1. Harbor City may therefore get more value from staggering crew handoffs and testing hub workers daily than from reducing passenger capacity on every route. The trade-off is data and complexity. A network-aware model can target interventions precisely, but it depends on roster data, route data, and assumptions that go stale if behavior changes faster than the model is updated.

Concept 3: Mutations and waves come from changing parameters in a changing graph

Halfway through Harbor City's response, sequencing finds a new norovirus lineage on the container-pier route. It does not create an entirely different disease model. It changes parameters inside the same model: higher transmissibility, partial immune escape, or a shorter generation interval. That small shift matters because transmission compounds. A variant that is 20 percent better at converting contacts into infections can become dominant quickly if it starts in the right part of the network.

Waves appear when several feedback loops line up. Early in the outbreak, many susceptible riders and dense commuter traffic make growth easy. Later, temporary immunity, school holidays, and cleaner vessels slow transmission. Then service normalizes, vigilance drops, and a fitter lineage arrives through the container route. The visible result is a second wave, but the mechanism is a sequence of parameter changes layered on top of the network: contact rate, susceptibility, variant fitness, and response behavior all moved.

That is why serious epidemic modeling uses scenario ranges instead of one frozen forecast. Harbor City should ask, "What happens if crew mixing falls by 30 percent?" and "What if the new lineage cuts the generation time by one day?" rather than pretending there is one authoritative curve. The trade-off is communication. Decision-makers often want a single number, while a responsible model produces a distribution of plausible paths. Still, that uncertainty is useful. It tells the transit authority which levers actually bend the outbreak and which interventions only look decisive on a dashboard.

The same mechanism also prepares the ground for the next lesson. Once disease dynamics alter staffing, rider demand, and inventory scarcity, decentralized price signals and market responses start to emerge on top of the outbreak itself.

Troubleshooting

Issue: The model matches the first week of cases and then misses the next wave completely.

Why it happens / is confusing: The transmission rate was treated as fixed even though Harbor City changed cleaning policy, schools reopened, and a new lineage entered through the container route.

Clarification / Fix: Refit the model with time-varying parameters or explicit intervention events. In epidemic work, a good fit to one regime does not guarantee the same parameters still apply in the next regime.

Issue: Closing one route seems to have little effect in the simulation even though operators know it is a high-risk corridor.

Why it happens / is confusing: The model is using geographic adjacency instead of actual contact structure. Crew swaps, maintenance overlap, and transfer behavior are missing, so the real bridge nodes do not look central.

Clarification / Fix: Build the network from observed contacts or rosters, not just from route maps. In Harbor City, the maintenance yard may matter more than the line that looks busiest on a tourist map.

Issue: Leaders celebrate because reported cases are falling, but workforce absences spike a few days later.

Why it happens / is confusing: Reported cases lag infections, and symptoms lag exposure. The operational burden is often driven by people who were infected before the dashboard turned downward.

Clarification / Fix: Track latent and infectious populations separately and pair case reports with leading indicators such as test positivity among crews, wastewater signals, or recent exposure estimates at the downtown terminal.

Advanced Connections

Connection 1: Epidemic Modeling <-> Opinion Dynamics

The previous lesson modeled how beliefs spread across Harbor City's fleet. Epidemic modeling uses many of the same graph ideas, but the state transition is biological rather than interpretive. Real incident response usually contains both layers at once: the virus spreads through contacts, and beliefs about risk spread through radio reports, dashboards, and policy meetings.

Connection 2: Epidemic Modeling <-> Market Dynamics

When Harbor City cuts routes, overtime prices rise, disinfectant inventory tightens, and riders shift demand to buses or remote work. Those responses are not side effects outside the model. They are another decentralized system reacting to scarcity and incentives. That is the bridge into 11.md, where the focus shifts from contagion curves to prices emerging from many local decisions.

Resources

Optional Deepening Resources

Key Insights

  1. Compartments turn counts into mechanism - Susceptible, exposed, infectious, and recovered states explain why curves move when they do.
  2. Networks decide where averages fail - Hubs, bridge nodes, and repeated contacts can dominate transmission even when citywide metrics look calm.
  3. Waves are feedback, not fate - Variants, behavior, immunity, and interventions keep changing the same underlying model parameters over time.
PREVIOUS Opinion Dynamics - The Physics of Belief NEXT Market Dynamics - Prices Emerge From Interactions

← Back to Agent-Based Modeling

← Back to Learning Hub