Skip to main content
Back

IIT's Neetyabhas Trains AI to Handle Messy Pandemic Data and Non-Compliance

Neetyabhas embeds real-world uncertainty and individual behavior into a hierarchical RL framework that optimizes pandemic interventions without assuming perfect data or compliance.

Paper, Full Access
  • New framework: Neetyabhas uses hierarchical RL to optimize pandemic interventions while accounting for imperfect data and policy non-compliance.
  • Two-tier simulation: 1,000 individual agents make mask, vaccine, and shopping decisions; a policymaker agent controls lockdowns and mandates.
  • Uncertainty injection: Eight algorithm variants (DDPG/TD3 with noise in actions, states, or both) are benchmarked; NR_DDPG scores highest at ~18 performance metrics.
  • Key finding: Masks and vaccines together most effectively reduced outbreak peak and duration; uncertainty-aware optimizers outperform vanilla baselines.
  • Behavioral realism: Panic-buying behavior in the simulation produced secondary epidemic waves, matching real COVID-19 dynamics.
  • Limitations: Only 1,000 agents simulated; no population heterogeneity; no public code release yet.

Most epidemic control models are built on a convenient fiction: that governments know exactly how many people are infected, and that when a lockdown is ordered, everyone complies. Neetyabhas, a new framework from researchers at IIT and collaborating institutions, tears down both assumptions at once. The result is a simulation that looks a lot more like the messy reality of pandemic response -- and an RL-driven policy optimizer that can handle it.

The gap between models and reality

Standard epidemic models treat the world as a math problem: plug in infection rates, output optimal lockdown schedules. Ordinary differential equation (ODE) models can optimize control, but the results don't translate easily to highly complex real-life environments. Agent-based methods allow detailed modeling but optimization is precluded by the large number of parameters. The deeper issue is that both approaches typically assume clean data and perfect policy execution.

In the real world, neither holds. Infection counts are undercounts. Hospitalization data lags. And when a government mandates masks or closes businesses, compliance is partial at best. While in a traditional MDP, the actions and the rewards are considered certain, in practice the directives of public policy -- such as the exact lockdown start and vaccination rollout -- cannot be followed perfectly, and the measurement of states such as the number of infected is also not precise. Neetyabhas is built specifically to model and optimize under these conditions.

A simulation with two layers of agents

The framework runs a population of 1,000 simulated individuals alongside a policymaker agent, all operating simultaneously. The disease itself is modeled using a 9-compartment epidemiological model -- a more granular extension of the classic SEIR framework (Susceptible, Exposed, Infected, Recovered) that adds states like Asymptomatic, Pre-symptomatic, Mild Infection, Severe Infection, Hospitalized, and Deceased. This level of granularity matters because it captures the distinct transmission dynamics of asymptomatic spreaders, a key driver of COVID-19's early waves.