comocma.como.Sofomore(interfaces.OOOptimizer)
class documentationcomocma.como
(View In Hierarchy)
Sofomore framework for multiobjective optimization, with the ask-and-tell interface. See: [Toure, Cheikh, et al. "Uncrowded Hypervolume Improvement: COMO-CMA-ES and the Sofomore framework." GECCO'19-Genetic and Evolutionary Computation Conference. 2019.]. Calling Sequences ================= - ``moes = Sofomore(list_of_solvers_instances, reference_point, opts)`` - ``moes = Sofomore(list_of_solvers_instances, reference_point)`` - ``moes = Sofomore(list_of_solvers_instances, reference_point).optimize(objective_fcts)`` Arguments ========= `list_of_solvers_instances` list of instances of single-objective solvers. It's generally created via a factory function. Let's take the example of a factory function that returns cma-es instances, called `get_cmas`. Then: ``list_of_solvers_instances = get_cmas(11 * [x0], sigma0)`` creates a list of 11 cma-es instances of initial mean `x0` and initial step-size `sigma0`. A single-objective solver instance must have the following attributes and methods: - `incumbent`: an attribute that gives an estimate of the single-objective solver. - `objective_values`: an attribute that stores the objective values of the incumbent. - `stop`: a method returning a dictionary representing the termination status. - `ask`: generates new candidate solutions. - `tell`: passes the objective values and updates the states of the single-objective solvers. `opts` opts, a dictionary with optional settings related to the Sofomore framework. It mainly contains the following keys: - 'archive': default value is `True`. If its value is `True`, tracks the non-dominated points that dominate the reference point, among the points evaluated so far during the optimization. The archive will not interfere with the optimization process. - 'indicator_front': default value is `None`. Used via `self.indicator_front = IndicatorFront(self.opts['indicator_front'])` within the __init__ method of Sofomore. See the class IndicatorFront for more details. - 'restart': used to define in __init__ the attribute `restart` via `self.restart = self.opts['restart']`. - 'update_order': default value is a function that takes a natural integer as input and returns a random number between 0 and 1. It is used as a `key value` in: `sorted(..., key = ...)`, and guides the order in which the kernels will be updated during the optimization. `reference_point` reference point of the multiobjective optimization. Its default value is `None` but should be set by the user beforehand to guarantee an optimal p-distribution convergence of the Hypervolume indicator of `p` points towards the Pareto set/front. It can be changed dynamically by the user if needed. Main interface / usage ====================== The interface is inherited from the generic `OOOptimizer` class, which is the same interface used by the `pycma` module. An object instance is generated as following:: >>> import cma, comocma >>> import numpy as np >>> reference_point = [11, 11] >>> num_kernels = 11 # the number of points we seek to have on the Pareto front >>> dimension = 10 # the dimension of the search space >>> x0 = dimension * [0] # an initial mean for cma >>> sigma0 = 0.2 # initial step-size for cma >>> list_of_solvers_instances = comocma.get_cmas(num_kernels * [x0], sigma0, {'verbose':-9}) >>> # `comocma.get_cmas` is a factory function that returns `num_kernels` cma-es instances >>> moes = comocma.Sofomore(list_of_solvers_instances, ... reference_point) #instantiation of our MO optimizer The least verbose interface is via the optimize method:: >>> fitness = comocma.FitFun(cma.ff.sphere, lambda x: cma.ff.sphere(x-1)) # a callable bi-objective function >>> moes.optimize(fitness) # doctest:+ELLIPSIS Iterat #Fevals Hypervolume axis ratios sigmas min&max stds*** More verbosely, the optimization of the callable multiobjective function `fitness` is done via the `ask-and-tell` interface:: >>> moes = comocma.Sofomore(list_of_solvers_instances, reference_point) >>> while not moes.stop() and moes.countiter < 30: ... solutions = moes.ask() # `ask` delivers new candidate solutions ... objective_values = [fitness(x) for x in solutions] ... moes.tell(solutions, objective_values) ... # `tell` updates the MO instance by passing the respective function values. ... moes.disp() # display data on the evolution of the optimization One iteration of the `optimize` interface is equivalent to one step in the loop of the `ask-and-tell` interface. But for the latter, the prototyper has more controls to analyse and guide the optimization, due to the access of the instance between the `ask` and the `tell` calls. Attributes and Properties ========================= - `archive`: list of non-dominated points among all points evaluated so far, which dominate the reference point. - `countevals`: the number of function evaluations. - `countiter`: the number of iterations. - `dimension`: is the dimension of the search space. - `indicator_front`: the indicator used as a changing fitness inside an iteration of Sofomore. By default it is the UHVI of all the objective values of the kernels' incumbents, except the kernel being optimized. See the class `IndicatorFront` for more details. - `isarchive`: a boolean accessible via `self.opts['archive']`. If `True`, we keep track of the archive, otherwise we don't. - `kernels`: initialized with `list_of_solvers_instances`, and is the list of single-objective solvers. - `key_sort_indices`: default value is `self.opts['update_order']`. It is a function used as a key to sort some kernels' indices, in order to select the first indices during the call of the `ask` method. - `logger`: an attribute that accounts for the way we log the Sofomore data during an optimization. `self.logger` is an instance of the SofomoreDataLogger class. - `NDA`: it is the non-dominated archiving method used within Sofomore. By default the class `BiobjectiveNondominatedSortedList` is used in two objectives and the class (non tested yet) `NonDominatedList` is used for three or more objectives. - `opts`: passed options. - `offspring`: list of tuples of the index of a kernel with its corresponding candidate solutions, that we generally get with the cma's `ask` method. - `pareto_front_cut`: list of non-dominated points among the incumbents of `self.kernels`, which dominate the reference point. - `pareto_set_cut`: preimage of `pareto_front_cut`. - `reference_point`: the current reference point. - `restart`: accessible via `self.opts['restart']`. If not `None`, a callback function that adds kernels after a kernel of the running Sofomore instance stops. Its default value is `None`, meaning that we do not do any restart. - `_active_indices`: the indices of the kernels that have not stopped yet. - `_last_stopped_kernel_id`: the index of the last stopped kernel. - `_told_indices`: the kernels' indices for which we will evaluate the objective values of their incumbents, before the next call of the `tell` method. And before the first call of `tell`, they are the indices of all the initial kernels (i.e. `range(len(self))`). And before another call of `tell`, they are the indices of the kernels from which we have sampled new candidate solutions during the penultimate `ask` method. Note that we should call the `ask` method before any call of the `tell` method.
Method | __init__ | No summary |
Method | __iter__ | make self iterable. |
Method | __getitem__ | make self subscriptable. |
Method | __len__ | return length of the Sofomore instance by calling len(.). |
Method | sorted | return a reversed sorted list of kernels. |
Method | ask | No summary |
Method | tell | No summary |
Method | pareto_front_cut | return the non-dominated solutions dominating the reference point, among the kernels' objective values. It's the image of self.pareto_set_cut . |
Method | pareto_set_cut | return the non-dominated solutions whose images dominate the reference point, among the kernels' incumbents. It's the pre-image of self.pareto_front_cut . |
Method | pareto_front_uncut | provisorial, return _all_ non-dominated objective values irrespectively |
Method | stop | No summary |
Method | termination_status | return a dictionary of the current termination states of the kernels. |
Method | add | add kernels of type list to self.kernels . Generally, kernels are created from a factory function. If kernels is of length 1, the brackets can be omitted. |
Method | remove | remove elements of the kernels (type list ) that belong to self.kernels , and update the _active_indices attribute. If kernels is of length 1, the brackets can be omitted. |
Method | median_stds | Undocumented |
Method | max_max_stds | Undocumented |
Method | inactivate | No summary |
Method | activate | activate kernel when it was inactivated beforehand. Otherwise it remains quiet. |
Method | disp_annotation | copy-pasted from cma.evolution_strategy . print annotation line for disp () |
Method | disp | copy-pasted from cma.evolution_strategy . print current state variables in a single-line. copy-pasted from cma.evolution_strategy module |
Method | _UHVI_indicator_archive | return archive for uncrowded hypervolume improvement indicator for kernel . |
Method | _UHVI_indicator | return indicator function(!) for uncrowded hypervolume improvement for kernel . |
Method | _indices_to_ask | Undocumented |
Initialization: - `list_of_solvers_instances` is a list of single-objective solvers' instances - The reference_point is set by the user during the initialization. - `opts` is a dictionary updating the values of 'indicator_front', 'archive', 'restart', 'update_order'; that respond respectfully to the changing fitness we will choose within an iteration of Sofomore, whether or not keeping an archive, how to do the restart in case of any restart, and the order of update of the kernels. It also has the keys 'verb_filename', 'verb_log' and 'verb_disp'; that respectfully indicate the name of the filename containing the Sofomore data, the logging of the Sofomore data every 'verb_log' iterations and the display of the data via `self.disp()` every 'verb_disp' iterations.
return length of the Sofomore
instance by calling len(.).
The length is the number of (active and inactive) kernels and hence consistent with subscription like [moes[i] for i in range(len(moes)) if i in moes._active_indices].
return archive for uncrowded hypervolume improvement indicator for kernel
.
kernel
can also be the respective index in self
.
return indicator function(!) for uncrowded hypervolume improvement for kernel
.
>>> import comocma, cma >>> list_of_solvers_instances = comocma.get_cmas(13 * [5 * [1]], 0.7, {'verbose':-9}) >>> fitness = comocma.FitFun(cma.ff.sphere, lambda x: cma.ff.sphere(x-1)) >>> moes = comocma.Sofomore(list_of_solvers_instances, [11, 11]) >>> moes.optimize(fitness, iterations=37) # doctest:+ELLIPSIS Iterat #Fevals Hypervolume axis ratios sigmas min&max stds*** >>> moes._UHVI_indicator(moes[1])(moes[2].objective_values) # doctest:+ELLIPSIS *** >>> moes._UHVI_indicator(1)(moes[2].objective_values) # doctest:+ELLIPSIS ***
both return the UHVI indicator function for kernel 1 and evaluate kernel 2 on it:
>>> [[moes._UHVI_indicator(k)(k.objective_values)] for k in moes] # doctest:+ELLIPSIS ***
is the list of UHVI values for all kernels where kernels occupying the very same objective value have indicator value zero.
return a reversed sorted list of kernels.
By default kernels are reversed sorted by HV contribution or UHVI (which we aim to maximize) in the set of kernels. Exact copies have zero or negative UHVI value.
>>> import comocma, cma >>> list_of_solvers_instances = comocma.get_cmas(13 * [5 * [1]], 0.7, {'verbose':-9}) >>> fitness = comocma.FitFun(cma.ff.sphere, lambda x: cma.ff.sphere(x-1)) >>> moes = comocma.Sofomore(list_of_solvers_instances, [11, 11]) >>> moes.optimize(fitness, iterations=31) # doctest:+ELLIPSIS Iterat #Fevals Hypervolume axis ratios sigmas min&max stds*** >>> moes.sorted(key = lambda k: moes.archive.contributing_hypervolume( ... k.objective_values)) # doctest:+ELLIPSIS [<comocma.como.CmaKernel object at***
sorts w.r.t. archive contribution (clones may get positive contribution).
get the kernels' incumbents to be evaluated and sample new candidate solutions from `number_to_ask` kernels. The sampling is done by calling the `ask` method of the `cma.CMAEvolutionStrategy` class. The indices of the considered kernels' incumbents are given by the `_told_indices` attribute. To get the `number_to_ask` kernels, we use the function `self.key_sort_indices` as a key to sort `self._remaining_indices_to_ask` (which is the list of kernels' indices wherein we choose the first `number_to_ask` elements. And if `number_to_ask` is larger than `len(self._remaining_indices_to_ask)`, we select the list `self._remaining_indices_to_ask` extended with the first `number_to_ask - len(self._remaining_indices_to_ask)` elements of `range(len(self))`, sorted with `self.key_sort_indices` as key. Arguments --------- - `number_to_ask`: the number of kernels for which we sample solutions from, it is of TYPE int or str and is smaller or equal to `len(self)`. The unique case where it is a str instance is when it is equal to "all", meaning that all the kernels are asked. That case is mainly used on parallel mode by distributing the evaluations of the kernels at the same time before calling the `tell` method. The opposite is when `number_to_ask` is equal to 1, which is the exact COMO-CMA-ES algorithm, where the evaluations are done sequentially by updating the kernels one by one. Return ------ The list of the kernels' incumbents to be evaluated, extended with a list of N-dimensional (N is the dimension of the search space) candidate solutions generated from `number_to_ask` kernels to be evaluated. :See: the `ask` method from the class `cma.CMAEvolutionStrategy`, in `evolution_strategy.py` from the `cma` module.
pass objective function values to update the state variables of some kernels, `self._told_indices` and eventually `self.archive`. Arguments --------- `solutions` list or array of points (of type `numpy.ndarray`), most presumably before delivered by the `ask()` method. `objective_values` list of multiobjective function values (of type `list`) corresponding to the respective points in `solutions`. `constraints_values` list of list of constraint values: each element is a list containing the values of one constraint function, that are obtained by evaluation on `solutions`. Details ------- To update a kernel, `tell()` applies the kernel's `tell` method to the kernel's corresponding candidate solutions (offspring) along with the "changing" fitness `- self.indicator_front.hypervolume_improvement`. :See: - the `tell` method from the class `cma.CMAEvolutionStrategy`, in `evolution_strategy.py` from the `cma` module. - the `hypervolume_improvement` method from the class `BiobjectiveNondominatedSortedList`, in the module `moarchiving.py` - the `hypervolume_improvement` method from the class `NonDominatedList`, in the module `nondominatedarchive.py`
self.pareto_set_cut
.self.pareto_front_cut
.provisorial, return _all_ non-dominated objective values irrespectively
of the current reference point. Only points whose contributing HV
does not depend on the current reference point still have the same
cHV in the resulting list
. HV improvements in general may change.
return a nonempty dictionary when all kernels stop, containing all the termination status. Therefore it's solely ... on the kernels' `stop` method, which also return dictionaries. Return ------ For example with 5 kernels, stop should return either None, or a `dict` of the form: {0: dict0, 1: dict1, 2: dict2, 3: dict2, 4: dict4}, where each index `i` is a key which value is the `dict` instance `self.kernels[i].stop()`
kernels
of type list
to self.kernels
.
Generally, kernels
are created from a factory function.
If kernels
is of length 1, the brackets can be omitted.kernels
(type list
) that belong to
self.kernels
, and update the _active_indices
attribute.
If kernels
is of length 1, the brackets can be omitted.kernel
, assuming that it's an element of self.kernels
,
or an index in range(len(self))
.
When inactivated, kernel
is no longer updated, it is ignored.
However we do not remove it from self.kernels
, meaning that kernel
might still play a role, due to its eventual trace in self.pareto_front_cut
.activate kernel
when it was inactivated beforehand. Otherwise
it remains quiet.
We expect the kernel's stop
method in interest to look like:
kernel.stop() = {'callback': ['kernel turned off']}
copy-pasted from cma.evolution_strategy
.
print current state variables in a single-line.
copy-pasted from cma.evolution_strategy
module
Prints only if iteration_counter % modulo == 0.
See Also | disp_annotation . |