Class: CrossEntropy::MatrixProblem

Inherits:
AbstractProblem show all
Defined in:
lib/cross_entropy/matrix_problem.rb

Overview

Assuming that the data are probabilities in an NArray (say dim 1 or dim 2 for now). Rows (NArray dimension 1) must sum to one. Columns (NArray dimension 0) represent the quantities to be optimized.

Caller should set seed with NArray.srand before calling.

Instance Attribute Summary

Attributes inherited from AbstractProblem

#elite_score, #max_iters, #min_score, #num_elite, #num_iters, #num_samples, #overall_min_score, #overall_min_score_sample, #params, #track_overall_min

Instance Method Summary collapse

Methods inherited from AbstractProblem

#for_stop_decision, #solve, #to_estimate, #to_generate_samples, #to_score_sample, #to_update

Constructor Details

#initialize(params = nil) {|_self| ... } ⇒ MatrixProblem

Returns a new instance of MatrixProblem.

Yields:

  • (_self)

Yield Parameters:



14
15
16
17
18
19
20
21
 
# File 'lib/cross_entropy/matrix_problem.rb', line 14

def initialize(params = nil)
  super(params)

  to_generate_samples { generate_samples_directly }
  to_estimate { |elite| estimate_ml(elite) }

  yield(self) if block_given?
end

Instance Method Details

#estimate_ml(elite) ⇒ NArray

Maximum likelihood estimate using only the given ‘elite’ solutions.

This is often (but not always) the optimal estimate for the probabilities from the elite samples for problems of this form.

on the AbstractProblem#num_elite parameter, but is typically less than AbstractProblem#num_samples; elements are integer in [0, #num_values)

non-negative floats in [0,1] and sum to 1

Parameters:

  • elite (NArray) —

    #num_variables rows; the number of columns depends

Returns:



55
56
57
58
59
60
61
62
63
64
65
 
# File 'lib/cross_entropy/matrix_problem.rb', line 55

def estimate_ml(elite)
  pr_est = NArray.float(num_values, num_variables)
  (0...num_variables).each do |i|
    elite_i = elite[true, i]
    (0...num_values).each do |j|
      pr_est[j, i] = elite_i.eq(j).count_true
    end
  end
  pr_est /= elite.shape[0]
  pr_est
end

#generate_samples_directly ⇒ Object

Generate samples directly from the probabilities matrix AbstractProblem#params.

If your problem is tightly constrained, you may want to provide a custom sample generation routine that avoids infeasible solutions; see AbstractProblem#to_generate_samples.



38
39
40
 
# File 'lib/cross_entropy/matrix_problem.rb', line 38

def generate_samples_directly
  params.tile(1, 1, num_samples).sample_pmf_dim.transpose(1, 0)
end

#most_likely_solution(pr = params) ⇒ Narray

Find most likely solution so far based on given probabilities.

#num_values columns; if not specified, the current AbstractProblem#params matrix is used

[0, #num_values)

Parameters:

  • pr (NArray) (defaults to: params) —

    probability matrix with #num_variables rows and

Returns:



77
78
79
80
81
82
83
84
 
# File 'lib/cross_entropy/matrix_problem.rb', line 77

def most_likely_solution(pr = params)
  pr_eq = pr.eq(pr.max(0).tile(1, pr.shape[0]).transpose(1, 0))
  pr_ml = NArray.int(pr_eq.shape[1])
  (0...pr_eq.shape[1]).each do |i|
    pr_ml[i] = pr_eq[true, i].where[0]
  end
  pr_ml
end

#num_values ⇒ Object 



27
28
29
 
# File 'lib/cross_entropy/matrix_problem.rb', line 27

def num_values
  @params.shape[0]
end

#num_variables ⇒ Object 



23
24
25
 
# File 'lib/cross_entropy/matrix_problem.rb', line 23

def num_variables
  @params.shape[1]
end