Merge pull request opencog#10 from ngeiswei/oc-to-sn

Merge opencog/moses to singnet/moses

moses/moses/main/moses_exec_def.h

@@ -36,7 +36,7 @@ using namespace std;
 // Available abbreviations are: O
 static const pair<string, string> rand_seed_opt("random-seed", "r");
 static const pair<string, string> problem_opt("problem", "H");
-static const pair<string, string> nsamples_opt("nsamples", "b");
+static const string nsamples_opt("nsamples");
 static const pair<string, string> min_rand_input_opt("min-rand-input", "q");
 static const pair<string, string> max_rand_input_opt("max-rand-input", "w");
 static const pair<string, string> max_evals_opt("max-evals", "m");

moses/moses/main/problem-params.cc

@@ -224,7 +224,7 @@ problem_params::add_options(boost::program_options::options_description& desc)
          "maj, demo, majority problem\n\n"
          "sr, demo, regression of f_n(x) = sum_{k=1,n} x^k\n")
 
-        (opt_desc_str(nsamples_opt).c_str(),
+        (nsamples_opt.c_str(),
          po::value<int>(&nsamples)->default_value(-1),
          "Number of samples to describe the problem. "
          "If nsample is negative, null or larger than the maximum "

moses/moses/man/moses.1

@@ -66,8 +66,7 @@ moses \- meta-optimizing semantic evolutionary search solver
 .IR algo ]
 .RB [ \-B
 .IR knob_effort ]
-.RB [ \-b
-.IR nsamples ]
+.RB [ \-\-nsamples ]
 .RB [ \-C1 ]
 .RB [ \-c
 .IR result_count ]
@@ -481,7 +480,7 @@ For boolean problems, this is the same as flipping the output
 value.  This option can only be used once, and, if used, it should
 specify a column containing an integer or floating-point value.
 .TP
-.BI \-b\  num \fR,\ \fB\-\-nsamples= num
+.BI \-\-nsamples= num
 The number of samples to be taken from the input file. Valid values
 run between 1 and the number of rows in the data file; other values
 are ignored. If this option is absent, then all data rows are used.
@@ -1249,7 +1248,7 @@ combo program specified with the \fB-y\fR flag. That is, the goal of
 the run is to deduce and learn the specified combo program.
 
 When specifying combo programs with continuous variables in them, be
-sure to use the \fB\-q\fR, \fB\-w\fR and \fB\-b\fR flags to specify
+sure to use \fB\-q\fR, \fB\-w\fR and \fB\-\-nsample\fR to specify
 a range of input values to be sampled. In order to determine the fitness
 of any candidate, it must be compared to the specified combo
 program.  The comparison is done at a variety of different input

moses/moses/optimization/hill-climbing.cc

@@ -28,6 +28,7 @@
  */
 
 #include <math.h>   // for sqrtf, cbrtf
+#include <algorithm>
 
 #include <boost/algorithm/minmax_element.hpp>
 
@@ -165,6 +166,7 @@ void hill_climbing::operator()(deme_t& deme,
 
             number_of_new_instances =
                 crossover(deme, current_number_of_instances,
+                          number_of_new_instances,
                           prev_start, prev_size, prev_center);
 
             already_xover = true;
@@ -532,11 +534,12 @@ size_t hill_climbing::n_new_instances(size_t distance, unsigned max_evals,
 // disabled and it leads to some massive slow down because then most
 // of the computational power is spent on successive representation
 // building
-#define MINIMUM_DEME_SIZE         100
+#define MINIMUM_DEME_SIZE         100UL
 
-    // If fraction is small, just use up the rest of the cycles.
+    // If fraction is small, just use up the rest of the cycles up to
+    // MINIMUM_DEME_SIZE.
     if (number_of_new_instances < MINIMUM_DEME_SIZE)
-        number_of_new_instances = nleft;
+        number_of_new_instances = std::min(nleft, MINIMUM_DEME_SIZE);
 
     if (nleft < number_of_new_instances)
         number_of_new_instances = nleft;
@@ -691,23 +694,29 @@ size_t hill_climbing::cross_top_three(deme_t& deme,
 }
 
 size_t hill_climbing::crossover(deme_t& deme, size_t deme_size,
+                                size_t max_number_of_new_instances,
                                 size_t sample_start, size_t sample_size,
                                 const instance& base) {
+    // Do not make more than the budget
+    size_t num_to_make = std::min(max_number_of_new_instances,
+                                  (size_t)hc_params.crossover_pop_size);
+
     // These cross-over (in the genetic sense) the
     // top-scoring one, two and three instances,respectively.
     size_t number_of_new_instances =
-        cross_top_one(deme, deme_size, hc_params.crossover_pop_size / 3,
+        cross_top_one(deme, deme_size,
+                      // Add the remainder so that it all sums up to
+                      // num_to_make in the end.
+                      num_to_make / 3 + num_to_make % 3,
                       sample_start, sample_size, base);
 
     number_of_new_instances +=
         cross_top_two(deme, deme_size + number_of_new_instances,
-                      hc_params.crossover_pop_size / 3,
-                      sample_start, sample_size, base);
+                      num_to_make / 3, sample_start, sample_size, base);
 
     number_of_new_instances +=
         cross_top_three(deme, deme_size + number_of_new_instances,
-                        hc_params.crossover_pop_size / 3,
-                        sample_start, sample_size, base);
+                        num_to_make / 3, sample_start, sample_size, base);
 
     return number_of_new_instances;
 }

moses/moses/optimization/hill-climbing.h

@@ -247,8 +247,11 @@ struct hill_climbing : optimizer_base
                            size_t sample_size,
                            const instance& base);
 
-    // chain the 3 crossovers methods above and return the number of new instances
+    // Chain the 3 crossovers methods above and return the number of
+    // new instances. The number of new instances should be lower or
+    // equal to max_number_of_new_instance.
     size_t crossover(deme_t& deme, size_t deme_size,
+                     size_t max_number_of_new_instance,
                      size_t sample_start, size_t sample_size,
                      const instance& base);
 

tests/moses/AntUTest.cxxtest

@@ -80,7 +80,7 @@ public:
         ant_bscore bscorer;
 
         // See the diary for the complexity ratio.
-        double complexity_ratio = 0.16;
+        double complexity_ratio = 0.08;
         bscorer.set_complexity_coef(complexity_ratio);
         behave_cscore cscorer(bscorer);