Evolution/Coevolution Labette

---

People

Sevan Greg Richard Pablo Hugues Paul
---

Projects

  • SEAM, for nearly-decomposable problems.
  • GAVIN, testing different crossover and mutation operators.
  • Pursuer-Evader coevolution with communication.
  • Tron, learning to play the video game - have a game! This will also be useful for education.
  • Evolutionary robotics, trying to coevolve a robot's body and mind, before building it.
  • Sprirals using massively parallel GP.
  • Learning Backgammon using speciation as categorical modularization.
  • Themes

  • Escalating arms race is the ideal.
  • Collusion: Stalls an arms race, Backgammon seems to suppress collusion.
  • Red Queen: Optimizing to a moving target makes it non-trivial to measure progress.
  • Difficult to make a finely-graded fitness function, coevolution can do this so you only have to learn what you almost already know, i.e., keeps the upgrade path in lots of small steps.
  • Speciation/niching: a way to avoid over-specialization and brittleness.
  • ---

    back to DEMO Private

    ---

    DEMO