enopt

Ensemble optimisation algorithm.

EnOpt

Bases: Optimize

This is an implementation of the ensemble steepest descent ensemble optimization algorithm - EnOpt. The update of the control variable is done with the simple steepest (or gradient) descent algorithm:

\[ x_l = x_{l-1} - \alpha \times C \times G \]

where \(x\) is the control variable, \(l\) is the iteration index, \(\alpha\) is the step size, \(C\) is a smoothing matrix (e.g., covariance matrix for \(x\)), and \(G\) is the ensemble gradient.

Methods:

Name	Description
calc_update	Update using steepest descent method with ensemble gradient

References

Chen et al., 2009, 'Efficient Ensemble-Based Closed-Loop Production Optimization', SPE Journal, 14 (4): 634-645.

TODO: Implement getter for optimize_result

__init__(fun, x, args, jac, hess, bounds=None, **options)

Parameters:

Name	Type	Description	Default
fun	callable	objective function	required
x	ndarray	Initial state	required
args	tuple	Initial covariance	required
jac	callable	Gradient function	required
hess	callable	Hessian function	required
bounds	list	(min, max) pairs for each element in x. None is used to specify no bound.	None
options	dict	Optimization options maxiter: maximum number of iterations (default 10) restart: restart optimization from a restart file (default false) restartsave: save a restart file after each successful iteration (defalut false) tol: convergence tolerance for the objective function (default 1e-6) alpha: step size for the steepest decent method (default 0.1) beta: momentum coefficient for running accelerated optimization (default 0.0) alpha_maxiter: maximum number of backtracing trials (default 5) resample: number indicating how many times resampling is tried if no improvement is found optimizer: 'GA' (gradient accent) or Adam (default 'GA') nesterov: use Nesterov acceleration if true (default false) hessian: use Hessian approximation (if the algorithm permits use of Hessian) (default false) normalize: normalize the gradient if true (default true) cov_factor: factor used to shrink the covariance for each resampling trial (defalut 0.5) savedata: specify which class variables to save to the result files (state, objective function value, iteration number, number of function evaluations, and number of gradient evaluations, are always saved)	{}

calc_update()

Update using steepest descent method with ensemble gradients