Skip to content

flowersteam/EpidemiOptim

master
Switch branches/tags

Name already in use

A tag already exists with the provided branch name. Many Git commands accept both tag and branch names, so creating this branch may cause unexpected behavior. Are you sure you want to create this branch?
Code

Latest commit

 

Git stats

Files

Permalink
Failed to load latest commit information.
Type
Name
Latest commit message
Commit time
 
 
 
 
 
 
 
 
 
 
 
 
 
 

The EpidemiOptim library

Left: evolution of costs and model parameters for a DQN lock-down policy DQN. Right: Pareto fronts found by 10 runs of NSGA-II.

The EpidemicOptimization suite is a library that allows to interface optimization algorithms with models of epidemic propagation. Epidemiological models are wrapped in OpenAI Gym interfaces, making them readily compatible with state-of-the-art optimization algorithms. The EpidemiOptim paper motivates the need for the EpidemiOptim library, describes it and presents a case study: the optimization of an on/off lockdown policy in the context of the COVID-19 pandemic.

This library is designed to be easily customized. The underlying propagation model, reward function, state and action spaces can be modified easily.

Click here to play online with the model

What's already there

  • Epidemiological models: This library comes with a region-based mechanistic model of the French COVID-19 epidemic fitted on real data, published here.

  • Cost functions: we include a health cost measuring the death toll, and an economic cost measuring the opportunity cost on the GDP due to the epidemic (see our paper for details). A multi-cost class allows to integrate the two costs.

  • Learning environments: The 'EpidemicDiscrete-v0' environment wraps around the epidemiological model and the two cost functions described above. The agent controls the enforcement of a lock-down on a weekly basis.

  • Optimization algorithms: The library includes NSGA-II, Deep Q-Networks (DQN), and variants of DQN presented in our paper.

Requirements

The EpidemiOptim library runs with Python>=3.6. All requirements are listed in the requirements.txt file.

Running an experiment

Experiments can be run with the epidemioptim/train.py script:

python train.py --config goal_dqn --expe-name goal_dqn_study --trial_id 0

The --config argument selects one of the configurations defined in epidemioptim/configs/. They include: ['dqn', 'goal_dqn', 'goal_dqn_constraints', 'nsga_ii']. The --expe_name and --trial_id arguments are a string identifier of the exeperiment and an integer identifier of the specific run that will be used to organized results in the data/results/ folder.

How to customize it

  • Add a module: epidemiological models, cost functions, gym environments or optimization algorithms can be added to the library. To this end, a new python script containing the module class should be created in epidemioptim/environments/models/, epidemioptim/environments/cost_functions/ ,epidemioptim/environments/gym_envs/ or epidemioptim/optimization/ respectively. They should inherit the corresponding base classes BaseModel, BaseCost, BaseEnv or BaseAlgorithm respectively. Finally, their reference should be added to their respective get functions (e.g. epidemioptim/environments/models/get_model.py).

Visualization

A visualization notebook can be found in epidemioptim/analysis/Visualization EpidemiOptim.ipynb. It allows the interactive visualizations of one run per condition presented in the paper. The user will be able to tune the parameter balance the economic and health costs, to modify constraints on their maximal cumulative values, or to explore the Pareto front resulting from the multi-objective optimization of the NSGA-II algorithm.

If you want to visualize your own runs, copy them from the data/results folder (where they are stored by the training process) to the data/data_for_visualization folder while respecting the organization.

This notebook is also hoster in an online application here. )

Please contribute !

The EpidemiOptim library can be seen as a collaborative toolbox to facilitate collaborations between people interested in the automatic design of intervention strategies in the context of epidemic propagation. Everyone is welcome to contribute. We are especially interested in new epidemiological models, new costs functions, new learning environments, new optimization algorithms and new ways to visualize the results. Feel free to contact us!

Please report any bug to us!

To cite the library:

@article{colas2020epidemioptim,
         title={EpidemiOptim: A Toolbox for the Optimization of Control Policies in Epidemiological Models},
         author={Colas, C{\'e}dric and Hejblum, Boris and Rouillon, S{\'e}bastien and Thi{\'e}baut, Rodolphe and Oudeyer, Pierre-Yves and Moulin-Frier,               Cl{\'e}ment and Prague, M{\'e}lanie},
         journal={arXiv preprint arXiv:2010.04452},
         year={2020}
}

About

Code for the paper EpidemiOptim: A Toolbox for the Optimization of Control Policies in Epidemiological Models.

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published