cma.constraints_handler.BoundPenalty

class documentation

class BoundPenalty(BoundaryHandlerBase):

Compute a bound penalty and update coordinate-wise penalty weights.

An instance must be updated each iteration using the update method.

Details:

The penalty computes like sum(w[i] * (x[i]-xfeas[i])**2), where xfeas is the closest feasible (in-bounds) solution from x. The weight w[i] should be updated during each iteration using the update method.

Example how this boundary handler is used with cma.fmin via the options (CMAOptions) of the class cma.CMAEvolutionStrategy:

>>> import cma
>>> res = cma.fmin(cma.ff.elli, 6 * [0.9], 0.1,
...     {'BoundaryHandler': cma.BoundPenalty,
...      'bounds': [-1, 1],
...      'tolflatfitness': 10,
...      'fixed_variables': {0: 0.012, 2:0.234}
...     })  # doctest: +ELLIPSIS
(4_w,8)-aCMA-ES (mu_w=2.6,w_1=52%) in dimension 4 (seed=...
>>> if res[1] >= 13.76: print(res)  # should never happen

Reference: Hansen et al 2009, A Method for Handling Uncertainty... IEEE TEC, with addendum, see https://ieeexplore.ieee.org/abstract/document/4634579 https://hal.inria.fr/inria-00276216/file/TEC2008.pdf

todo: implement a more generic interface, where this becomes a fitness wrapper which adds the desired penalty and the update method is used as callback argument for fmin like:

f = cma.BoundPenalty(cma.ff.elli, bounds=[-1, 1])
res = cma.fmin(f, 6 * [1], callback=f.update)

where callback functions should receive the same arguments as tell, namely an CMAEvolutionStrategy instance, an array of the current solutions and their respective f-values. Such change is relatively involved. Consider also that bounds are related with the geno- to phenotype transformation.

Method	`__call__`	returns the boundary violation penalty for `x`, where `x` is a single solution or a list or np.array of solutions.
Method	`__init__`	Argument bounds can be `None` or `bounds[0]` and `bounds[1]` are lower and upper domain boundaries, each is either `None` or a scalar or a `list` or `np.array` of appropriate size.
Method	`feasible_ratio`	counts for each coordinate the number of feasible values in `solutions` and returns an `np.array` of length `len(solutions[0])` with the ratios.
Method	`repair`	sets out-of-bounds components of `x` on the bounds.
Method	`update`	updates the weights for computing a boundary penalty.
Instance Variable	`gamma`	Undocumented
Instance Variable	`hist`	Undocumented
Instance Variable	`weights_initialized`	Undocumented

Inherited from BoundaryHandlerBase:

Method	`get_bound`	return lower and upper bound of variable with index `index`
Method	`get_bounds`	`get_bounds('lower', 8)` returns the lower bounds in 8-D
Method	`has_bounds`	return `True` if any variable is bounded
Method	`idx_out_of_bounds`	return index list of out-of-bound values in `x`.
Method	`inverse`	inverse of repair if it exists, at least it should hold `repair == repair o inverse o repair`
Method	`is_in_bounds`	not yet tested
Method	`to_dim_times_two`	return boundaries in format `[[lb0, ub0], [lb1, ub1], ...]`, as used by `BoxConstraints...` class.
Instance Variable	`bounds`	Undocumented
Method	`_get_bounds`	ib == 0/1 means lower/upper bound, return a vector of length `dimension`

def __call__(self, x, archive, gp): ¶

overrides cma.constraints_handler.BoundaryHandlerBase.__call__

returns the boundary violation penalty for x, where x is a single solution or a list or np.array of solutions.

def __init__(self, bounds=None): ¶

overrides cma.constraints_handler.BoundaryHandlerBase.__init__

Argument bounds can be None or bounds[0] and bounds[1] are lower and upper domain boundaries, each is either None or a scalar or a list or np.array of appropriate size.

def feasible_ratio(self, solutions): ¶

counts for each coordinate the number of feasible values in solutions and returns an np.array of length len(solutions[0]) with the ratios.

def repair(self, x, copy_if_changed=True): ¶

overrides cma.constraints_handler.BoundaryHandlerBase.repair

sets out-of-bounds components of x on the bounds.

def update(self, function_values, es): ¶

overrides cma.constraints_handler.BoundaryHandlerBase.update

updates the weights for computing a boundary penalty.

Arguments

function_values:: all function values of recent population of solutions
es:: CMAEvolutionStrategy object instance, in particular mean and variances and the methods from the attribute gp of type GenoPheno are used.

gamma = ¶

Undocumented

hist: list = ¶

Undocumented

weights_initialized: bool = ¶

Undocumented