year: 2024
paper: https://arxiv.org/abs/2410.02543
website:
code: https://github.com/RobertTLange/evosax/blob/main/evosax/algorithms/population_based/diffusion_evolution.py
connections: diffusion models, evolutionary algorithm, levin, particle swarm optimization, Benedikt Hartl, levin
“Unprecedented quality diversity scores.”
Todo
FOllow up paperd Heuristically Adaptive Diffusion-Model Evolutionary Strategy
thoughts while listening to the talk
the exploration landscape is so… incredibly interesting and diverse
which well makes sense because you start with random noise over the space
why don’t other algos do this. this seems to smart and natural
natural in the sense that nature also has like a lot of branches it explores in parallel
whihc niching in other algos tries to do, as you can see with MAP-Elites
→ diversity
→ multiple good solutions found
→ more efficient sampling
→ easier to overcome local optima
https://discord.com/channels/1184542508156469310/1412730295911841854/1412732844811485206