cadCAD_machine_search

cadCAD tools for preparing & analyzing experiments where large-scale machine search for selecting parameters are involved.

How the choice of parameters that optimizes a given KPI affects the combined KPIs

Installation

The recommended way is through pip:

pip install cadCAD_machine_search

Alternatively, you can install a branch directly from the repository through:

pip install git+https://github.com/danlessa/cadCAD_machine_search@master

Usage

Tools

Cartesian product for parameter sweeping

# The sweep_cartesian_product makes it easy to create large-scale simulations
from cadcad_machine_search.tools import sweep_cartesian_product

PARAMS_TO_SWEEP = {    
    "predator_death_const": np.linspace(0.9, 1.1, 5),
    "prey_death_const": np.linspace(0.02, 0.04, 5),
    "dt": np.linspace(0.04, 0.06, 4)
    }

sweep_params = sweep_cartesian_product(PARAMS_TO_SWEEP)

params = {
    "prey_birth_rate": [1.0],
    "predator_birth_rate": [0.01],
    **sweep_params

Visualizations

Sensitivity of a KPI towards a set of control parameters

from cadcad_machine_search.visualizations import kpi_sensitivity_plot

# KPI: prey population is 5x above the predator populaton at least 90% of the time
control_params = set(sweep_params.keys())
WEIGHT = 5.0
kpi = lambda df: (df.prey_population > WEIGHT * df.predator_population).mean() > 0.9
kpi_sensitivity_plot(df, kpi, control_params)

Impact of parameter selection in terms of specific vs general goals

from cadcad_machine_search.visualizations import plot_goal_ternary

PREY_ABUNDANCE_THRESHOLD = (df.prey_population / df.predator_population).mean()
PREDATOR_ABUNDANCE_THRESHOLD = (df.predator_population / df.prey_population).mean()
PREY_VARIANCE_THRESHOLD = df.prey_population.std()
PREDATOR_VARIANCE_THRESHOLD = df.predator_population.std()

kpis = {'prey_abundance': lambda df: df.prey_population.mean(),
        'predator_abundance': lambda df: df.predator_population.mean(),
        'prey_variance': lambda df: 1 / df.prey_population.std(),
        'predator_variance': lambda df: 1 / df.predator_population.std(),
        }

# Define goals as sides of the triangle
goals = {}
goals['Prey Desirability'] = lambda metrics: (metrics['prey_abundance'] + metrics['prey_variance']) / 2
goals['Predator Desirability'] = lambda metrics: (metrics['predator_abundance'] + metrics['predator_variance']) / 2
goals['System Stability'] = lambda metrics: (metrics['prey_variance'] + metrics['predator_variance']) / 2
goals['combined'] = lambda goals: goals[0] + goals[1] + goals[2]

plot_goal_ternary(df, kpis, goals, control_params)

Examples

We recommend having a look on the notebooks/ folder, as we have a P&P model that makes usage of the parameter sweep helpers together with the KPI sensitivity visualization.