Back to browse
Watch a neural net discover molecules by arguing with itself

Watch a neural net discover molecules by arguing with itself

by cmaicurious·May 14, 2026·1 point·0 comments

AI Analysis

●●●BangerWizardryRabbit HoleBig Brain

Two RL clones duel in-browser with epiplexity selection — no human tuning the explore-exploit knob.

Strengths
  • Epiplexity-based winner selection (critic learning gain) is a genuinely novel meta-RL approach.
  • Runs entirely in-browser with TF.js and RDKit — no server, no GPU, real training visible live.
  • Sparse-reward molecular discovery with SELFIES representation and Tanimoto similarity metrics.
Weaknesses
  • Educational demo rather than production tool — limited to ~75k params and simple targets.
  • Niche audience: only ML folks who care about RL architecture will dig into the mechanics.
Category
Target Audience

ML researchers, students learning reinforcement learning

Similar To

TensorFlow.js RL demos · ChemOS · GuacaMol

Post Description

Uses epiplexity selection and meta-rl to find molecules

Similar Projects