optimization.asktell
Gradient-free tuning of Params by ask and tell: you run the episodes, it picks the next.
Functions
bounds_of(params, names)
|
Lower and upper bounds and current values of the named Params as flat vectors. |
tune(searcher, evaluate, rounds)
|
rounds times: ask, evaluate each candidate with evaluate(x), tell.
|
Classes
Bayes(lower, upper[, initial, pool, seed])
|
Bayesian optimization: a Gaussian process of the costs so far, and the next candidate where it expects the most improvement. |
CMAES(x0, sigma[, lower, upper, size, seed])
|
(mu/mu_w, lambda)-CMA-ES from the mean x0 with step size sigma; default lambda is 4 + floor(3 ln n). |
ExtremumSeeking(x0, amplitude, gain[, ...])
|
Gradient descent from two runs per round: the point is perturbed by amplitude up and down along a random direction, and moves gain times the slope those runs show, downhill. |
Grid(lower, upper, points)
|
Every point of a regular grid between the bounds, asked once; points per dimension. |
Random(lower, upper[, size, seed])
|
size uniform samples inside the bounds at every ask.
|
