Skip to content

Latest commit

 

History

History
64 lines (53 loc) · 3.12 KB

ARCHITECTURE.md

File metadata and controls

64 lines (53 loc) · 3.12 KB
Raw

Architecture

This file provides a simple explanation for the design and purpose of Chromoo's classes and how they couple. This ends up being a description of how pymoo is designed, since we are necessarily restricted to implement our program under the constraints of the optimization framework.

Below, you will find a rough overview of the classes used in Chromoo, how they couple together to allow running the optimization process, and some general features of this program.

Classes

  • ChromooProblem -> pymoo.core.problem.Problem
    • has _evaluate() to do run simulations for every generation
  • ChromooCallback -> pymoo.core.callback.Callback
    • Is run after every generation
    • Used to update cache and store population and Pareto front data
  • pymoo.util.termination.default.MultiObjectiveDefaultTermination
    • Termination criteria for the optimizer
  • AlgorithmFactory -> Use pymoo.{algorithms,factory,operators}
    • Build and init an algorithm.
  • CadetSimulation -> cadet.Cadet
    • Extensions of the existing Cadet class
    • Allow loading config as yaml
    • Methods to calculate solute mass in each domain
  • Objective
    • Calculate objective scores given a completed cadet simulation
    • Can handle slicing/averaging/summing of multi-dimensional solution data
  • Parameter
    • Container for a given parameter (or list)
    • Can handle scalar, vector, or element (specific indices in a vector) as parameter
  • Scores
    • module to store scoring functions
    • sse, rmse, nrmse, logsse, logrmse
  • Transforms
    • module to store transform (normalization) functions
    • lognorm, lognorm_inverse, normalize, denormalize, identity

Architecture / Design / Coupling

  • Algorithm setup loads the Problem, Termination criteria, Callback, and Initial Population
  • Algorithms can be run step-by-step (step = generation)
  • On each run, a callback can be provided to perform specific tasks after each step.
    • We use this to save data into a pd.Dataframe via our Cache class
  • Problems have a _evaluate method that does the work. It takes the parameters as an array, and returns an array of objective scores.
    • The inputs X are provided by pymoo, and may be in the normalized space. Thus a denormalization may be required before passing them on to the actual evaluation process.
    • In our case, each individual in the given population is evaluated in its own thread (Set nthreads=1 in the cadet simulation template), and multiple evaluations can be performed in parallel.

Features

Fitting

  • normalization of data for parameters and objectives

  • sobol initialization

  • restart fitting from last generation

  • store population results for all generations

  • store latest Pareto front

  • pymoo v0.5 algorithms: UNSGA3/NSGA3

  • Fit scalars, full vectors, or specific elements of vectors

  • Allow slicing high-dimensional vectors to generate objectives

Post

  • Convergence plots (Metascore/objectivescores per generation)
  • Response plots (Parameter vs Objective Score)
  • Performance plots (Objectives vs Reference)
  • Corner plots (High dimensional views of specific dataset)
  • Violin plots for parameter values in Pareto front