Skip to content

initial_proposal.md

Latest commit

 

History

History
69 lines (52 loc) · 2.44 KB

initial_proposal.md

File metadata and controls

69 lines (52 loc) · 2.44 KB

Initial Proposal for a DecisionVis MOEA

Algorithmic Elements

We propose a MOEA using the following well-established algorithmic elements:

  • steady-state (EA Handbook 262)
  • tournament selection (EA Handbook 256)
  • UM (EA Handbook 285)
  • SBX and PM (Deb '99)
  • ε-dominance archiving (Laumans et al 2002 (proposal), Deb 2003 (εMOEA)
  • restarts (Coello '01 μGA)
  • adaptive population sizing (Kollat '05 εNSGAII)
  • nondomination sorting (Deb 2002 NSGAII)

Asynchronous Implementation

This implementation will be asynchronous, meaning that evolution consists of two main procedures: generate individual and receive evaluated individual.

The generate individual procedure produces a new population member for evaluation.

The receive evaluated individual procedure receives an evaluated individual and updates the state of the population accordingly.

Notably, the asynchronous approach relies on an external "main loop" to manage the generation, evaluation, and return of population members. This allows users a great deal of latitude to:

  • Return population members that were not generated by the MOEA.
    • Enables the reuse of solutions produced by previous optimization runs.
    • Enables the addition of local search to improve solutions
  • Determine when and how the MOEA state will be captured.
  • Pause and continue optimization runs.
  • Parallelize the evaluation of population members (if model evaluation is expensive).

Python Language

The Python programming language is

  • highly productive
  • broadly supported
  • widely adopted by the scientific community
  • well known to the implementors

License

The proposed MOEA will be licensed and released to the public under the GNU General Public License (version 3).

DecisionVis, as the copyright holder, will reserve the right to incorporate this MOEA in its own proprietary offerings, as well as to license this MOEA to others under different terms.

Comment

We expect that this dual licensing situation will discourage reciprocal contributions from outside the company, although under the terms of the GPL others are free to take the algorithm and modify it for their own purposes. Our goal in adopting these terms is to put out a simple, modern, easy-to-use MOEA and thereby encourage adoption. We've tried selling MOEAs, but nobody's buying. Instead we hope that broad adoption, particularly with our name attached, will bring us customers with interesting problems to collaborate on.