odow / SDDP.jl

A JuMP extension for Stochastic Dual Dynamic Programming
https://sddp.dev
Other
308 stars 61 forks source link

Clarify meaning of first column in output #113

Closed lkapelevich closed 6 years ago

lkapelevich commented 6 years ago

I don't think there are updated docs clarifying this at the moment. Just leaving this here to answer a question.

The first column is either a single value or a confidence interval from a Monte Carlo simulation. When the entry is a single value, its is the objective from a single (and the most recent) forward pass in SDDP. The single value is not an average.

To compute a statistical upper bound, you can: 1) Use the simulate function after you finished solving your model 2) If you want to get bounds while solving, use the simulation keyword argument in the solve function. You can control how frequently these will run.

There is an example of both in: https://github.com/odow/SDDP.jl/blob/master/examples/Newsvendor/newsvendor.jl.

The first time we solve the model, we set the frequency of simulations to be every 10 iterations: https://github.com/odow/SDDP.jl/blob/master/examples/Newsvendor/newsvendor.jl#L99 So in the 10th iteration you will see a calculation of % Gap and a confidence interval.

Then there is an example using the simulate function: https://github.com/odow/SDDP.jl/blob/master/examples/Newsvendor/newsvendor.jl#L134 And in the line below we compute the mean objective of 500 simulations.

odow commented 6 years ago

This was addressed partially in https://github.com/odow/SDDP.jl/pull/112. The docs should update in the next few minutes.

odow commented 6 years ago

The new docs address this: https://odow.github.io/SDDP.jl/latest/tutorial/02_rhs_noise.html#Solving-the-problem-1