presp

Open-source ESP implementation for Project Resilience

Evolution

The evolutionary step assumes it is given a sorted, evaluated population.

1. (Optional) Elites are taken from the population and used to update the evaluator (full ESP loop).
2. The next population is created using crossover/mutation of the current population.
3. The next population is evaluated and sorted.

Usage

The Evolution class is the core of the library and takes in many evolution parameters as well as the Evaluator and the PrescriptorFactory.

3 things have to be implemented to use the library: the Evaluator, the Prescriptor, and the PrescriptorFactory.

• The Prescriptor is an individual in the population whose forward function must be implemented for use the in the Evaluator.
• The PrescriptorFactory must implement crossover/mutation for evolution.
• The Evaluator takes in a population of Prescriptors and attaches metrics to each individual.

A basic example would look like this:

from presp.evolution import Evolution

from examples.evaluator import ImplementedEvaluator
from examples.prescriptor import ImplementedFactory

config = {...}
evaluator = ImplementedEvaluator(**config["eval_params"])
prescriptor_factory = ImplementedFactory(**config["prescriptor_params"])

evolution = Evolution(**config["evolution_params"], evaluator=evaluator, prescriptor_factory=prescriptor_factory)

evolution.run_evolution()

Example

For a more full-fledged example see examples/cartpole for a simple implementation of direct evolution (not ESP) on the CartPole gymnasium environment.

See examples/bnh for a simple example of directly evolving the parameters of the BNH function.