Base definition of algorithms¶
- class orion.algo.base.BaseAlgorithm(space, **kwargs)[source]¶
Base class describing what an algorithm can do.
- Parameters
- space
orion.algo.space.Space
Definition of a problem’s parameter space.
- kwargsdict
Tunable elements of a particular algorithm, a dictionary from hyperparameter names to values.
- space
Notes
We are using the No Free Lunch theorem’s [R367dceedad15-1]_[R367dceedad15-3]_ formulation of an
BaseAlgorithm
. We treat it as a part of a procedure which in each iteration suggests a sample of the parameter space of the problem as a candidate solution and observes the results of its evaluation.Developer Note: Each algorithm’s complete specification, i.e. implementation of its methods and parameters of its own, lies in a separate concrete algorithm class, which must be an immediate subclass of
BaseAlgorithm
. [The reason for this is current implementation oforion.core.utils.Factory
metaclass which uses BaseAlgorithm.__subclasses__().] Second, one must declare an algorithm’s own parameters (tunable elements which could be set by configuration). This is done by passing them to BaseAlgorithm.__init__() by calling Python’s super with a Space object as a positional argument plus algorithm’s own parameters as keyword arguments. The keys of the keyword arguments passed to BaseAlgorithm.__init__() are interpreted as the algorithm’s parameter names. So for example, a subclass could be as simple as this (regarding the logistics, not an actual algorithm’s implementation):References
- 1
D. H. Wolpert and W. G. Macready, “No Free Lunch Theorems for Optimization,” IEEE Transactions on Evolutionary Computation, vol. 1, no. 1, pp. 67–82, Apr. 1997.
- 2
W. G. Macready and D. H. Wolpert, “What Makes An Optimization Problem Hard?,” Complexity, vol. 1, no. 5, pp. 40–46, 1996.
- 3
D. H. Wolpert and W. G. Macready, “No Free Lunch Theorems for Search,” Technical Report SFI-TR-95-02-010, Santa Fe Institute, 1995.
Examples
1from orion.algo.base import BaseAlgorithm 2from orion.algo.space import (Integer, Space) 3 4class MySimpleAlgo(BaseAlgorithm): 5 6 def __init__(self, space, multiplier=1, another_param="a string param"): 7 super().__init__(space, multiplier=multiplier, another_param=another_param) 8 9 def suggest(self, num=1): 10 print(self.another_param) 11 return list(map(lambda x: tuple(map(lambda y: self.multiplier * y, x)), 12 self.space.sample(num))) 13 14 def observe(self, points, results): 15 pass 16 17dim = Integer('named_param', 'norm', 3, 2, shape=(2, 3)) 18s = Space() 19s.register(dim) 20 21algo = MySimpleAlgo(s, 2, "I am just sampling!") 22algo.suggest()
- Attributes
configuration
Return tunable elements of this algorithm in a dictionary form appropriate for saving.
fidelity_index
Returns the name of the first fidelity dimension if there is one, otherwise None.
has_completed_max_trials
Returns True if the algorithm has a max_trials attribute, and has completed more trials than its value.
is_done
Whether the algorithm is done and will not make further suggestions.
n_observed
Number of completed trials observed by the algorithm.
n_suggested
Number of trials suggested by the algorithm
- requires_dist
- requires_shape
- requires_type
space
Domain of problem associated with this algorithm’s instance.
state_dict
Return a state dict that can be used to reset the state of the algorithm.
Methods
get_id
(trial[, ignore_fidelity, ignore_parent])Return unique hash for a trials based on params
has_observed
(trial)Whether the algorithm has observed a given point objective.
has_suggested
(trial)Whether the algorithm has suggested a given point.
Returns True if the algorithm has more trials in its registry than the number of possible values in the search space.
judge
(trial, measurements)Inform an algorithm about online measurements of a running trial.
observe
(trials)Observe the results of the evaluation of the trials in the process defined in user's script.
register
(trial)Save the trial as one suggested or observed by the algorithm.
score
(trial)Allow algorithm to evaluate point based on a prediction about this parameter set's performance.
seed_rng
(seed)Seed the state of the random number generator.
set_state
(state_dict)Reset the state of the algorithm based on the given state_dict
should_suspend
(trial)Allow algorithm to decide whether a particular running trial is still worth to complete its evaluation, based on information provided by the
judge
method.suggest
(num)Suggest a num of new sets of parameters.
- property configuration¶
Return tunable elements of this algorithm in a dictionary form appropriate for saving.
- property fidelity_index¶
Returns the name of the first fidelity dimension if there is one, otherwise None.
- get_id(trial, ignore_fidelity=False, ignore_parent=False)[source]¶
Return unique hash for a trials based on params
The trial is assumed to be in the transformed space if the algorithm is working in a transformed space.
- Parameters
- trialTrial
trial from a
orion.algo.space.Space
.- ignore_fidelity: bool, optional
If True, the fidelity dimension is ignored when computing a unique hash for the trial. Defaults to False.
- ignore_parent: bool, optional
If True, the parent id is ignored when computing a unique hash for the trial. Defaults to False.
- property has_completed_max_trials: bool¶
Returns True if the algorithm has a max_trials attribute, and has completed more trials than its value.
- has_observed(trial)[source]¶
Whether the algorithm has observed a given point objective.
This only counts observed completed trials.
- Parameters
- trial: ``orion.core.worker.trial.Trial``
Trial object to retrieve from the database
- Returns
- bool
True if the trial’s objective was observed by the algo, False otherwise.
- has_suggested(trial)[source]¶
Whether the algorithm has suggested a given point.
- Parameters
- trial: ``orion.core.worker.trial.Trial``
Trial from a
orion.algo.space.Space
.
- Returns
- bool
True if the trial was suggested by the algo, False otherwise.
- has_suggested_all_possible_values() bool [source]¶
Returns True if the algorithm has more trials in its registry than the number of possible values in the search space.
If there is a fidelity dimension in the search space, only the trials with the maximum fidelity value are counted.
- property is_done: bool¶
Whether the algorithm is done and will not make further suggestions.
Return True, if an algorithm holds that there can be no further improvement. By default, the cardinality of the specified search space will be used to check if all possible sets of parameters has been tried.
- judge(trial, measurements)[source]¶
Inform an algorithm about online measurements of a running trial.
This method is to be used as a callback in a client-server communication between user’s script and a orion’s worker using a
BaseAlgorithm
. Data returned from this method must be serializable and will be used as a response to the running environment. Default response is None.- Parameters
- trial: ``orion.core.worker.trial.Trial``
Trial object to retrieve from the database
- Returns
- None or a serializable dictionary containing named data
Notes
Calling algorithm to
judge
a point based on its online measurements will effectively change a state in the algorithm (like a reinforcement learning agent’s hidden state or an automatic early stopping mechanism’s regression), which it may change the value of the propertyshould_suspend
.
- property n_observed¶
Number of completed trials observed by the algorithm.
- property n_suggested¶
Number of trials suggested by the algorithm
- observe(trials)[source]¶
Observe the results of the evaluation of the trials in the process defined in user’s script.
- Parameters
- trials: list of ``orion.core.worker.trial.Trial``
Trials from a
orion.algo.space.Space
.
- register(trial)[source]¶
Save the trial as one suggested or observed by the algorithm.
- Parameters
- trial: ``orion.core.worker.trial.Trial``
a Trial from self.space.
- score(trial)[source]¶
Allow algorithm to evaluate point based on a prediction about this parameter set’s performance.
By default, return the same score any parameter (no preference).
- Parameters
- trial: ``orion.core.worker.trial.Trial``
Trial object to retrieve from the database
- Returns
- A subjective measure of expected performance.
- seed_rng(seed)[source]¶
Seed the state of the random number generator.
- Parameters
seed – Integer seed for the random number generator.
Note
This methods does nothing if the algorithm is deterministic.
- set_state(state_dict)[source]¶
Reset the state of the algorithm based on the given state_dict
- Parameters
state_dict – Dictionary representing state of an algorithm
- should_suspend(trial)[source]¶
Allow algorithm to decide whether a particular running trial is still worth to complete its evaluation, based on information provided by the
judge
method.
- property space¶
Domain of problem associated with this algorithm’s instance.
- property state_dict¶
Return a state dict that can be used to reset the state of the algorithm.
- abstract suggest(num: int) list[Trial] [source]¶
Suggest a num of new sets of parameters.
- Parameters
- num: int
Number of points to suggest. The algorithm may return less than the number of points requested.
- Returns
- list of trials or None
A list of trials representing values suggested by the algorithm. The algorithm may opt out if it cannot make a good suggestion at the moment (it may be waiting for other trials to complete), in which case it will return None.
Notes
New parameters must be compliant with the problem’s domain
orion.algo.space.Space
.IMPORTANT: Algorithms must call self.register(trial) for every trial that is returned by this method. This is important for the algorithm to be able to keep track of the trials it has suggested/observed, and for the auto-generated unit-tests to pass.