sorry if duplicate.
i have two-class prediction model; has n configurable (numeric) parameters. model can work pretty if tune parameters properly, specific values parameters hard find. used grid search (providing, say, m values each parameter). yields m ^ n times learn, , very time-consuming when run in parallel on machine 24 cores.
i tried fixing parameters 1 , changing 1 parameter (which yields m × n times), it's not obvious me results got. sample plot of precision (triangles) , recall (dots) negative (red) , positive (blue) samples:
simply taking "winner" values each parameter obtained way , combining them doesn't lead best (or good) prediction results. thought building regression on parameter sets precision/recall dependent variable, don't think regression more 5 independent variables faster grid search scenario.
what propose find parameter values, reasonable estimation time? sorry if has obvious (or well-documented) answer.
i use randomized grid search (pick random values each of parameters in given range deem reasonable , evaluate each such randomly chosen configuration), can run long can afford to. this paper runs experiments show @ least grid search:
grid search , manual search used strategies hyper-parameter optimization. paper shows empirically , theoretically randomly chosen trials more efficient hyper-parameter optimization trials on grid. empirical evidence comes comparison large previous study used grid search , manual search configure neural networks , deep belief networks. compared neural networks configured pure grid search, find random search on same domain able find models or better within small fraction of computation time.
for it's worth, have used scikit-learn's random grid search problem required optimizing 10 hyper-parameters text classification task, results in around 1000 iterations.
Comments
Post a Comment