NAS Performance Estimation: Weight Sharing vs Full Training

Neural Architecture Search has to rank thousands of candidate networks it can never afford to fully train, so real NAS systems trade estimation accuracy for speed. This simulator builds a small weight-sharing supernet — six stages of six candidate operations each — and lets a search controller sample child architectures (paths through the graph) while you switch the performance-estimation strategy between Full Training, Early Stopping and Weight Sharing. Each strategy has its own simulated compute cost and its own gap between the fast proxy score and the ground-truth accuracy a path would get if trained standalone; a live Spearman-style correlation readout shows that gap actually widening as Weight Sharing's shared weights co-adapt to whichever paths get sampled most, exactly the effect documented in the NAS "one-shot" evaluation literature.