Parameter Space Exploration and CMA-ES Optimisation

This package might be useful if you have a model taking as input a set of parameters.

Start simulations from any sample(s) drown from the parameter space. Sampling may be analytical (like ParameterSpaceExploration with param iterators) or unknown (like ParameterSpaceAsk for CMA-ES).

The simulation parameters are the main model parameters. Optimisation parameters define a subset of simu params and spawn the parameter space. This package basically builds SimulationParameters for any sample of the parameter space, allowing to easily start simulations, and optimisation.

Install

pip install optimflow

Run parameters exploration

out_path = Path(__file__).parent / "out"

# Declare parameters of interest
class Params(OptimParams):
    p1 = param.Number(1, bounds=(0, 1), doc="Parameter 1")
    p2 = param.Number(1, bounds=(0.9, 1.1), doc="Parameter 2")

# Create simulation and exploration parameters
simulation_params = SimulationParams()
optim_params = Params()

# Define your simulation worker
def worker(dname: str):
    # Load parameters from the directory (prepared by dump_params)
    params = SimulationParams.load_from(dname)

    # mock the model's result
    t = np.linspace(0, 1, 100)
    res = params.p1 * np.sin(params.p2 * 2 * np.pi * t)
    np.savetxt(params.out_dir / "result.txt", res)

# Main class to start exploring the parameter space
explo = ParameterSpaceExploration(out_path)
explo.dump_params(simulation_params, optim_params)  # linear iterator
explo.run(worker, parallel=True)  # start all models
explo.gather_results()  # build a big pandas dataframe
explo.plot_results()  # plot model results for each varying optim param