Novelty Search

  • method used in genetic algorithms / fitness functions
    • instead of trying to optimise for single "best" solution, optimise for "new", previously unseen solutions
    • the idea being that evolution is not a product of optimising fitness, but rather a product of open-ended discovery
  • I find this interesting when applied to research (and maybe life even?)
    • stating the goals too rigidly or too quickly, and then comparing every solution against them (like a fitness function would) can lead to finding non-optimal solutions (also once a solution is found, the search can "stop" - which feels more like engineering-y approach than science-y approach)

The fundamental problem is that the stepping stones that lead to the objective may not resemble the objective itself.

— Novelty Search and the Problem with Objectives

Contrary to intuition, searching without regard to the objective can often outperform searching explicitly for the objective.

— Novelty Search and the Problem with Objectives

natural evolution succeeds because it divergently explores many ways of life while optimizing a behavior (i.e. reproduction) largely orthogonal to what is interesting about its discoveries, while objective-based search directly follows the gradient of improvement until it either succeeds or is too far deceived

— Novelty Search and the Problem with Objectives

I found out a couple of interesting things. First of all, magic lenses are fun to use, and feel distinctively different to what most of the computing is like right now. Because the transformations are pure functions (they don't destroy the underlying material, just like a physical lens wouldn't), you can experiment without any fear of losing the original data. You also often get unexpected results, which is great for Novelty Search.

  • one idea is applying Novelty Search meta-heuristic learnings to real life
    • don't worry about clearly defined final goal (because that goal then was probably stated by someone else, and it's the Hamming's thing of "important problems in your field" narrowing down the field and important problems for you)
    • focus on building your own niches

the evolution of mathematics, art, and technology are facilitated by exploration around recent discoveries, serendipity, and a plethora of diverse and conflicting individual objectives. That is, these human-driven processes of search also do not aim at any unified society-wide singular objective. Thus the types of search processes that continually innovate to produce radical advancements often lack a final predefined goal. This observation makes sense because a single fixed goal would either (1) be deceptive and therefore bring search to a point at which progress would effectively halt, or (2) if the goal is not so deceptive then innovation would cease once the goal is met.

— Novelty Search and the Problem with Objectives

innovation may result more from accumulating novel ways of life (i.e. new niches) than from optimizing fitness

— Novelty Search and the Problem with Objectives

Szymon Kaliski © 2021
mailtwittergithubwebring