31
Genetic algorithm learns how to make functional Factorio blueprints (youtu.be)
submitted 3 months ago* (last edited 3 months ago) by to c/factorio@lemmy.world
15 comments fedilink hide all child comments
 

This seems neat.

you are viewing a single comment's thread
view the rest of the comments
[–] 14 points 3 months ago (1 child)

"Generative AI" is its own type of ML typically thought of as modela that generate unstructured data (text, video, audio).

Genetic Algorithms don't typically fall into that category as they produce structured outputs. They work by starting with a handful of models, letting them run, then using the top performing models to generate new models - hence the genetic part.

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 2 child comments
    •
  • [–] 5 points 3 months ago (1 child)

    Doesn't the stuff typically considered Generative AI use genetic algorithms too, at least for training? Why would a Factorio blueprint optimized for performance be considered structured while text optimized to resemble other text is unstructured?

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 2 child comments
    •
  • [–] 5 points 3 months ago (1 child)

    I'm not aware of genetic algorithms being used in GenAI. Genetic algorithms really had their hayday in the 1990s to my understanding. They're generally very limited due to the heuristic like approach to refining the underlying model.

    As for structured vs unstructured, you can typically think of it as whether or not the information is "machine readable" or has an explicit structure.

    With the Factorio setting the structure is the grid (how many spaces are there to work with) and what structures you can place on each grid space. So in an x by y rectangle the algorithm has to decide what of the z types of structures should be at each spot on the grid (effectively an xyz output space)

    Conversley with language/text you run into many different problems. Just trying to define how many words there are is an ever evolving challenge. While you might be limited to 26 characters, there are no rules about how long or short a sentence is, grammer is flexible, conjugations are all over the place, and the meaning of a sentence can change based on the last word Garden path sentence. Essentially, there is no definitive meaning of any sentence without a contextual lens to provide it structure.

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 2 child comments
    •
  • [–] 3 points 3 months ago (1 child)

    Ok, so what you're saying is it's about whether the goal can be precisely defined.

    Afaik the basic way they are trained is conceptually very similar or the same as genetic algorithms: GANs are similar to mimicry in evolutionary biology, with an evolutionary arms race between both networks.

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 2 child comments
    •
  • [–] 5 points 3 months ago

    Haha, I'm happy to go deep on this if you want, but will keep this post shortish.

    Types of data (structured vs unstructured) and types of ML (supervised vs unsupervised vs reinforcement learning) are different.

    The way Generative AI models are trained is generally very different than genetic algorithms and typically uses a mixture of Reinforcement Learning and Supervised learning with a diffusion or transformer backend (type of deep learning layer).

    GANs probably count as generative AI, but predate the current AI models by a few years. Their approach is more about competing models than survival of the fittest, as it has a model that generates and a model that critiques. Their main drawback was that they could really only output content in a narrow range (faces or landscapes, but not both).

  • permalink
  • fedilink
  • source
  • parent