search

Yelp / MOE

A global, black box optimization engine for real world metric optimization.
Other
1.31k stars 139 forks source link

Grant user the ability to choose number of mc iterations for calculating ExpectedImprovement and gradient EI seperately #423

Open jialeiwang opened 9 years ago

jialeiwang commented 9 years ago

Motivation: While stochastic gradient descent algorithm has tolerance to the noise coming from monte carlo estimation of gradient, the noise from estimation of EI has great impact to the performance of the algorithm. To find a better balance between performance vs efficiency, we can set number of mc iterations for estimating gradient low, while number of mc iterations for estimating EI high.

suntzu86 commented 9 years ago

If you have some results demonstrating the value/trade-off, I would love to see them :)