SBI for Epidemic Parameter Inference on Adaptive Networks

This repository contains code, data, notebooks, and report material for simulation-based inference (SBI) on a stochastic adaptive-network SIR model.

Goal: infer unknown parameters $(\beta,\gamma,\rho)$ from aggregate observations when the likelihood is intractable.

Model summary

• Population: 200 agents on an undirected Erdos-Renyi contact network ($p=0.05$).
• States: Susceptible (S), Infected (I), Recovered (R).
• Initial condition: 5 infected agents at time 0.
• Time horizon: 200 steps.

At each step:

1. Infection: each S-I edge transmits with probability $\beta$.
2. Recovery: each infected node recovers with probability $\gamma$.
3. Rewiring: each S-I edge rewires away from infected contact with probability $\rho$.

Parameter priors used by ABC methods:

Parameter	Meaning	Prior
$\beta$	Infection probability per S-I edge per step	Uniform(0.05, 0.50)
$\gamma$	Recovery probability per infected agent per step	Uniform(0.02, 0.20)
$\rho$	Rewiring probability per S-I edge per step	Uniform(0.0, 0.8)

Project structure

Core code:

• simulator.py: adaptive-network SIR simulator.
• ABC_rejection.py: rejection ABC with pilot scaling and quantile acceptance.
• SMC_ABC.py: SMC-ABC with weighted particles, stage-wise tolerance tightening, and diagnostics.

Observed dataset:

• data/infected_timeseries.csv
• data/rewiring_timeseries.csv
• data/final_degree_histograms.csv

Analysis notebooks:

• parameter_sweep_visualization.ipynb: parameter-sweep diagnostics and refined summary-statistic selection.
• abc_rejection_analysis.ipynb: rejection ABC workflow, posterior plots, diagnostics.
• abc_smc_analysis.ipynb: SMC-ABC workflow, stage diagnostics, weighted posterior plots, rejection-vs-SMC comparison.

Setup

Recommended environment:

1. Create and activate a virtual environment.
2. Install dependencies:

pip install numpy pandas matplotlib jupyter

Quick usage

Run rejection ABC:

python ABC_rejection.py --n-pilot 100 --n-samples 2000 --accept-quantile 0.05 --n-reps 20 --out-prefix abc_rejection_analysis

Run SMC-ABC:

python SMC_ABC.py --n-pilot 100 --n-particles 100 --n-init-samples 1000 --n-stages 4 --stage-quantile 0.60 --kernel-scale 0.60 --n-reps 20 --out-prefix abc_smc_analysis

Open notebooks:

jupyter notebook

Summary statistics used in ABC runs

The refined 9-dimensional summary vector is:

• infected_extinction_t
• degree_std
• rewire_peak
• infected_auc
• degree_tail_low
• infected_final
• degree_tail_hi
• infected_peak
• infected_t_peak

These were selected from parameter-sweep diagnostics to balance:

1. sensitivity to parameter changes,
2. complementarity (low overlap),
3. joint informativeness for $(\beta,\gamma,\rho)$,
4. compact dimensionality.

Notes

• SMC-ABC generally improves efficiency and posterior concentration relative to basic rejection ABC in this project.
• A residual $\beta$ - $\rho$ trade-off remains and should be interpreted as a structural identifiability limitation rather than a pure sampling artifact.