Skip to content

Latest commit

 

History

History
106 lines (82 loc) · 6.4 KB

optimization.md

File metadata and controls

106 lines (82 loc) · 6.4 KB
Raw

Optimization

All optimizers assume a centered and reduced prior at the beginning of the optimization (i.e. 0 mean and unitary standard deviation). They are however able to find solutions far from this initial prior.

Basic example

Optimizing (minimizing!) a function using an optimizer (here OnePlusOne) can be easily run with:

import nevergrad as ng

def square(x, y=12):
    return sum((x - .5)**2) + abs(y)

optimizer = ng.optimizers.OnePlusOne(instrumentation=2, budget=100)
# alternatively, you could use ng.optimizers.registry["OnePlusOne"]
# (registry is a dict containing all optimizer classes)
recommendation = optimizer.minimize(square)
print(recommendation)
>>> Candidate(args=(array([0.500, 0.499]),), kwargs={})

recommendation holds the optimal attributes args and kwargs found by the optimizer for the provided function. In this example, the optimal value will be found in recommendation.args[0] and will be a np.ndarray of size 2.

instrumentation=n is a shortcut to state that the function has only one variable, of dimension n, Defining the following instrumentation instead will optimize on both x and y.

instrum = ng.Instrumentation(ng.var.Array(2), y=ng.var.Array(1).asscalar())
optimizer = ng.optimizers.OnePlusOne(instrumentation=instrum, budget=100)
recommendation = optimizer.minimize(square)
print(recommendation)
>>> Candidate(args=(array([0.490, 0.546]),), kwargs={'y': 0.0})

See the instrumentation tutorial for more complex instrumentations.

Using several workers

Running the function evaluation in parallel with several workers is as easy as providing an executor:

from concurrent import futures
optimizer = ng.optimizers.OnePlusOne(instrumentation=instrum, budget=100, num_workers=5)
with futures.ProcessPoolExecutor(max_workers=optimizer.num_workers) as executor:
    recommendation = optimizer.minimize(square, executor=executor, batch_mode=False)

With batch_mode=True it will ask the optimizer for num_workers points to evaluate, run the evaluations, then update the optimizer with the num_workers function outputs, and repeat until the budget is all spent. Since no executor is provided, the evaluations will be sequential. num_workers > 1 with no executor is therefore suboptimal but nonetheless useful for evaluation purpose (i.e. we simulate parallelism but have no actual parallelism). batch_mode=False (steady state mode) will ask for a new evaluation whenever a worker is ready.

Ask and tell interface

An ask and tell interface is also available. The 3 key methods for this interface are respectively:

  • ask: suggest a candidate on which to evaluate the function to optimize.
  • tell: for updated the optimizer with the value of the function for a candidate.
  • provide_recommendation: returns the candidate the algorithms considers the best. For most optimization algorithms in the platform, they can be called in arbitrary order - asynchronous optimization is OK. Some algorithms (with class attribute no_parallelization=True however do not support this.

The Candidate class holds attributes args and kwargs corresponding to the args and kwargs of the function you optimize, given its instrumentation. It also holds a data attribute corresponding to the data point in the optimization space.

Here is a simpler example in the sequential case (this is what happens in the optimize method for num_workers=1):

for _ in range(optimizer.budget):
    x = optimizer.ask()
    value = square(*x.args, **x.kwargs)
    optimizer.tell(x, value)
recommendation = optimizer.provide_recommendation()

Please make sure that your function returns a float, and that you indeed want to perform minimization and not maximization ;)

Choosing an optimizer

You can print the full list of optimizers with:

import nevergrad as ng
print(sorted(ng.optimizers.registry.keys()))

All algorithms have strengths and weaknesses. Questionable rules of thumb could be:

  • TwoPointsDE is excellent in many cases, including very high num_workers.
  • PortfolioDiscreteOnePlusOne is excellent in discrete settings of mixed settings when high precision on parameters is not relevant; it's possibly a good choice for hyperparameter choice.
  • OnePlusOne is a simple robust method for continuous parameters with num_workers < 8.
  • CMA is excellent for control (e.g. neurocontrol) when the environment is not very noisy (num_workers ~50 ok) and when the budget is large (e.g. 1000 x the dimension).
  • TBPSA is excellent for problems corrupted by noise, in particular overparameterized (neural) ones; very high num_workers ok).
  • PSO is excellent in terms of robustness, high num_workers ok.
  • ScrHammersleySearchPlusMiddlePoint is excellent for super parallel cases (fully one-shot, i.e. num_workers = budget included) or for very multimodal cases (such as some of our MLDA problems); don't use softmax with this optimizer.
  • RandomSearch is the classical random search baseline; don't use softmax with this optimizer.

Optimizing machine learning hyperparameters

When optimizing hyperparameters as e.g. in machine learning. If you don't know what variables (see instrumentation) to use:

  • use SoftmaxCategorical for discrete variables
  • use TwoPointsDE with num_workers equal to the number of workers available to you. See the machine learning example for more.

Or if you want something more aimed at robustly outperforming random search in highly parallel settings (one-shot):

  • use OrderedDiscrete for discrete variables, taking care that the default value is in the middle.
  • Use ScrHammersleySearchPlusMiddlePoint (PlusMiddlePoint only if you have continuous parameters or good default values for discrete parameters).

Reproducibility

Each instrumentation has its own random_state for generating random numbers. All optimizers pull from it when they require stochastic behaviors. For reproducibility, this random state can be seeded in two ways:

  • by setting numpy's global random state seed (np.random.seed(32)) before the instrumentation's first use. Indeed, when first used, the instrumentation's random state is seeded with a seed drawn from the global random state.
  • by manually seeding the instrumentation random state (E.g.: instrumentation.random_state.seed(12) or optimizer.instrumentation.random_state = np.random.RandomState(12))