🧪 2D REINVENT-Style Augmented-Likelihood Policy for Atom Chains: Prior, Target Hetero-Atom Fraction and the Gibbs Bound
Stacked policy bars for each chain position against the frozen prior, the latest generated atom chains with their scores, a REINVENT versus plain REINFORCE switch, and six independent numerical checks (exact enumeration, policy-gradient theorem, Gibbs identity).
⚙ Under the hood
Stacked policy bars for each chain position against the frozen prior, the latest generated atom chains with their scores, a REINVENT versus plain REINFORCE switch, and six independent numerical checks (exact enumeration, policy-gradient theorem, Gibbs identity).
REINVENTPolicy GradientReinforcement LearningGenerative ChemistryAugmented LikelihoodKL DivergenceGibbs Variational PrincipleAI Drug Discovery Infrastructure
2D · HTML5 Canvas 2D · 60 FPS target · runs fully client-side, no install