🗣️ 2D Emergent Signalling Game: REINFORCE Speaker and Listener Learning a Shared Convention
Front view of an emergent-communication signalling game: tabular softmax speaker and listener trained by REINFORCE, policy matrices, noisy-channel error, exact expected payoff and mutual information, checked against finite differences, a 60000-episode unbiasedness test and a brute-force count of the N! conventions.
⚙ Under the hood
Front view of an emergent-communication signalling game: tabular softmax speaker and listener trained by REINFORCE, policy matrices, noisy-channel error, exact expected payoff and mutual information, checked against finite differences, a 60000-episode unbiasedness test and a brute-force count of the N! conventions.
Emergent CommunicationSignalling GameREINFORCEPolicy GradientMulti-Agent LearningLewis ConventionMutual InformationNoisy Channel
2D · HTML5 Canvas 2D · 60 FPS target · runs fully client-side, no install