public class GamePlayerBenchmark extends java.lang.Object
This simulates game-player-like activity, with a few CoNLL CliqueTrees playing host to lots and lots of manipulations by adding and removing human "observations". In real life, this kind of behavior occurs during sampling lookahead for LENSE-like systems.
In order to measure only the realistic parts of behavior, and not the random generation of numbers, we pre-cache a few hundred ConcatVectors representing human obs features, then our feature function is just indexing into that cache. The cache is designed to require a bit of L1 cache eviction to page through, so that we don't see artificial speed gains during dot products b/c we already have both features and weights in L1 cache.
|Modifier and Type||Class and Description|
|Constructor and Description|
Stanford NLP Group