Population synthesis as scenario generation

1. Installation

To install the package, clone the repository and run:

pip install synthpop

2. Documentation

You can view the docs here. In particular, you will find examples for how to apply the methods contained in this package to generate populations and scenarios of interest in example agent-based simulators.

3. Example

Consider a population of $N$ agents whose states $x_i \sim \mathcal{N}(\mu_i, 1)$, where $\mathcal{N}(\mu, \sigma^2)$ denotes a Normal distribution with mean $\mu$ and variance $\sigma^2$. Consider also generating the agent-level attributes $\mu_i$ from a distribution $\iota_\mu = \mathcal{N}(\mu, 1)$. We'd like to find a proposal distribution $q$ over the population-level parameter $\mu$ such that the average square

$$\ell(\mathbf{x}) = \frac{1}{N} \sum_{i=1}^{N} x_i^2$$

of the agent states is small.

3.1 Implementing the model

We implement the model for how agent parameters $\mu_i$ are generated given $\mu$, along with the model for how the agent states $x_i$ are forward simulated given their individual $\mu_i$:

import numpy as np
import warnings
from ..abstract import AbstractModel

class Normals(AbstractModel):
    def __init__(self, n_timesteps=1, n_agents=1_000):
        self.n_timesteps = n_timesteps
        self.n_agents = n_agents

    def initialize(self):
        pass

    def step(self, *args, **kwargs):
        pass

    def observe(self, x):
        return [x]

    @staticmethod
    def make_default_generator(params):
        mu = params

        # Specify how the population parameter \mu parameterises the agent generator
        def generator(n_agents):
            # Draw agent parameters from distribution \iota_\mu
            mus = mu + np.random.normal(size=n_agents)
            return mus

        return generator

    def run(self, generator):
        with warnings.catch_warnings():
            warnings.simplefilter("ignore")
            # Generate agent parameters \mu_i
            mus = generator(self.n_agents)
            # Simulate model forward to obtain the x_i
            xs = mus + np.random.normal(size=self.n_agents)
            return self.observe(xs)

3.2 Specifying the loss function

We also specify the loss function:

import torch

def loss(x):
    z = torch.mean(torch.pow(x[0], 2))
    return z

3.3 Wrapping the agent attribute generator

We wrap this for convenience:

class AgentAttributeDistributionGenerator(SampleGenerator):
    def forward(self, generator_params):
        mu = generator_params
        return model.make_default_generator(mu)

meta_generator = AgentAttributeDistributionGenerator()

3.4 Specify the domain and optimise

Finally, we specify the domain over which we'd like to find such a $q$, and a method for obtaining $q$, before running the optimisation procedure:

prior = torch.distributions.Uniform(torch.tensor([-20.]), torch.tensor([20.]))

optimise = Optimise(model=model, meta_generator=meta_generator, prior=prior, loss=loss)
optimise_method = TBS_SMC(num_particles=5_000, num_initial_pop=10_000, num_simulations=10_000, epsilon_decay=0.7, return_summary=True)
trained_meta_generator = optimise.fit(optimise_method, num_workers=-1)

3.5 Optimising with variational optimisation

The same example, optimised using variational optimisation, can be seen here.

4. Citation

This package accompanies our AAMAS 2024 paper on Population synthesis as scenario generation in agent-based models, with the aim of facilitating simulation-based planning under uncertainty. You can cite our paper and/or package using the following:

@inproceedings{dyer2023a,
  publisher = {Association for Computing Machinery},
  title = {Population synthesis as scenario generation for simulation-based planning under uncertainty},
  author = {Dyer, J and Quera-Bofarull, A and Bishop, N and Farmer, JD and Calinescu, A and Wooldridge, M},
  year = {2023},
  organizer = {23rd International Conference on Autonomous Agents and Multiagent Systems (AAMAS 2024)},
}

The supplementary material (that being this GitHub repository and the paper appendix) can be cited separately from the main paper as:

@software{joel_dyer_2024_10629106,
  author       = {Joel Dyer and
                  Arnau Quera-Bofarull},
  title        = {joelnmdyer/synthpop: AAMAS release},
  month        = feb,
  year         = 2024,
  publisher    = {Zenodo},
  version      = {v1.0.0},
  doi          = {10.5281/zenodo.10629106},
  url          = {https://doi.org/10.5281/zenodo.10629106}
}