1 // Copyright 2017 The Abseil Authors.
2 //
3 // Licensed under the Apache License, Version 2.0 (the "License");
4 // you may not use this file except in compliance with the License.
5 // You may obtain a copy of the License at
6 //
7 // https://www.apache.org/licenses/LICENSE-2.0
8 //
9 // Unless required by applicable law or agreed to in writing, software
10 // distributed under the License is distributed on an "AS IS" BASIS,
11 // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12 // See the License for the specific language governing permissions and
13 // limitations under the License.
14
15 #ifndef ABSL_RANDOM_DISCRETE_DISTRIBUTION_H_
16 #define ABSL_RANDOM_DISCRETE_DISTRIBUTION_H_
17
18 #include <cassert>
19 #include <cstddef>
20 #include <initializer_list>
21 #include <istream>
22 #include <limits>
23 #include <ostream>
24 #include <type_traits>
25 #include <utility>
26 #include <vector>
27
28 #include "absl/base/config.h"
29 #include "absl/random/bernoulli_distribution.h"
30 #include "absl/random/internal/iostream_state_saver.h"
31 #include "absl/random/uniform_int_distribution.h"
32
33 namespace absl {
34 ABSL_NAMESPACE_BEGIN
35
36 // absl::discrete_distribution
37 //
38 // A discrete distribution produces random integers i, where 0 <= i < n
39 // distributed according to the discrete probability function:
40 //
41 // P(i|p0,...,pn−1)=pi
42 //
43 // This class is an implementation of discrete_distribution (see
44 // [rand.dist.samp.discrete]).
45 //
46 // The algorithm used is Walker's Aliasing algorithm, described in Knuth, Vol 2.
47 // absl::discrete_distribution takes O(N) time to precompute the probabilities
48 // (where N is the number of possible outcomes in the distribution) at
49 // construction, and then takes O(1) time for each variate generation. Many
50 // other implementations also take O(N) time to construct an ordered sequence of
51 // partial sums, plus O(log N) time per variate to binary search.
52 //
53 template <typename IntType = int>
54 class discrete_distribution {
55 public:
56 using result_type = IntType;
57
58 class param_type {
59 public:
60 using distribution_type = discrete_distribution;
61
param_type()62 param_type() { init(); }
63
64 template <typename InputIterator>
param_type(InputIterator begin,InputIterator end)65 explicit param_type(InputIterator begin, InputIterator end)
66 : p_(begin, end) {
67 init();
68 }
69
param_type(std::initializer_list<double> weights)70 explicit param_type(std::initializer_list<double> weights) : p_(weights) {
71 init();
72 }
73
74 template <class UnaryOperation>
param_type(size_t nw,double xmin,double xmax,UnaryOperation fw)75 explicit param_type(size_t nw, double xmin, double xmax,
76 UnaryOperation fw) {
77 if (nw > 0) {
78 p_.reserve(nw);
79 double delta = (xmax - xmin) / static_cast<double>(nw);
80 assert(delta > 0);
81 double t = delta * 0.5;
82 for (size_t i = 0; i < nw; ++i) {
83 p_.push_back(fw(xmin + i * delta + t));
84 }
85 }
86 init();
87 }
88
probabilities()89 const std::vector<double>& probabilities() const { return p_; }
n()90 size_t n() const { return p_.size() - 1; }
91
92 friend bool operator==(const param_type& a, const param_type& b) {
93 return a.probabilities() == b.probabilities();
94 }
95
96 friend bool operator!=(const param_type& a, const param_type& b) {
97 return !(a == b);
98 }
99
100 private:
101 friend class discrete_distribution;
102
103 void init();
104
105 std::vector<double> p_; // normalized probabilities
106 std::vector<std::pair<double, size_t>> q_; // (acceptance, alternate) pairs
107
108 static_assert(std::is_integral<result_type>::value,
109 "Class-template absl::discrete_distribution<> must be "
110 "parameterized using an integral type.");
111 };
112
discrete_distribution()113 discrete_distribution() : param_() {}
114
discrete_distribution(const param_type & p)115 explicit discrete_distribution(const param_type& p) : param_(p) {}
116
117 template <typename InputIterator>
discrete_distribution(InputIterator begin,InputIterator end)118 explicit discrete_distribution(InputIterator begin, InputIterator end)
119 : param_(begin, end) {}
120
discrete_distribution(std::initializer_list<double> weights)121 explicit discrete_distribution(std::initializer_list<double> weights)
122 : param_(weights) {}
123
124 template <class UnaryOperation>
discrete_distribution(size_t nw,double xmin,double xmax,UnaryOperation fw)125 explicit discrete_distribution(size_t nw, double xmin, double xmax,
126 UnaryOperation fw)
127 : param_(nw, xmin, xmax, std::move(fw)) {}
128
reset()129 void reset() {}
130
131 // generating functions
132 template <typename URBG>
operator()133 result_type operator()(URBG& g) { // NOLINT(runtime/references)
134 return (*this)(g, param_);
135 }
136
137 template <typename URBG>
138 result_type operator()(URBG& g, // NOLINT(runtime/references)
139 const param_type& p);
140
param()141 const param_type& param() const { return param_; }
param(const param_type & p)142 void param(const param_type& p) { param_ = p; }
143
result_type(min)144 result_type(min)() const { return 0; }
result_type(max)145 result_type(max)() const {
146 return static_cast<result_type>(param_.n());
147 } // inclusive
148
149 // NOTE [rand.dist.sample.discrete] returns a std::vector<double> not a
150 // const std::vector<double>&.
probabilities()151 const std::vector<double>& probabilities() const {
152 return param_.probabilities();
153 }
154
155 friend bool operator==(const discrete_distribution& a,
156 const discrete_distribution& b) {
157 return a.param_ == b.param_;
158 }
159 friend bool operator!=(const discrete_distribution& a,
160 const discrete_distribution& b) {
161 return a.param_ != b.param_;
162 }
163
164 private:
165 param_type param_;
166 };
167
168 // --------------------------------------------------------------------------
169 // Implementation details only below
170 // --------------------------------------------------------------------------
171
172 namespace random_internal {
173
174 // Using the vector `*probabilities`, whose values are the weights or
175 // probabilities of an element being selected, constructs the proportional
176 // probabilities used by the discrete distribution. `*probabilities` will be
177 // scaled, if necessary, so that its entries sum to a value sufficiently close
178 // to 1.0.
179 std::vector<std::pair<double, size_t>> InitDiscreteDistribution(
180 std::vector<double>* probabilities);
181
182 } // namespace random_internal
183
184 template <typename IntType>
init()185 void discrete_distribution<IntType>::param_type::init() {
186 if (p_.empty()) {
187 p_.push_back(1.0);
188 q_.emplace_back(1.0, 0);
189 } else {
190 assert(n() <= (std::numeric_limits<IntType>::max)());
191 q_ = random_internal::InitDiscreteDistribution(&p_);
192 }
193 }
194
195 template <typename IntType>
196 template <typename URBG>
197 typename discrete_distribution<IntType>::result_type
operator()198 discrete_distribution<IntType>::operator()(
199 URBG& g, // NOLINT(runtime/references)
200 const param_type& p) {
201 const auto idx = absl::uniform_int_distribution<result_type>(0, p.n())(g);
202 const auto& q = p.q_[idx];
203 const bool selected = absl::bernoulli_distribution(q.first)(g);
204 return selected ? idx : static_cast<result_type>(q.second);
205 }
206
207 template <typename CharT, typename Traits, typename IntType>
208 std::basic_ostream<CharT, Traits>& operator<<(
209 std::basic_ostream<CharT, Traits>& os, // NOLINT(runtime/references)
210 const discrete_distribution<IntType>& x) {
211 auto saver = random_internal::make_ostream_state_saver(os);
212 const auto& probabilities = x.param().probabilities();
213 os << probabilities.size();
214
215 os.precision(random_internal::stream_precision_helper<double>::kPrecision);
216 for (const auto& p : probabilities) {
217 os << os.fill() << p;
218 }
219 return os;
220 }
221
222 template <typename CharT, typename Traits, typename IntType>
223 std::basic_istream<CharT, Traits>& operator>>(
224 std::basic_istream<CharT, Traits>& is, // NOLINT(runtime/references)
225 discrete_distribution<IntType>& x) { // NOLINT(runtime/references)
226 using param_type = typename discrete_distribution<IntType>::param_type;
227 auto saver = random_internal::make_istream_state_saver(is);
228
229 size_t n;
230 std::vector<double> p;
231
232 is >> n;
233 if (is.fail()) return is;
234 if (n > 0) {
235 p.reserve(n);
236 for (IntType i = 0; i < n && !is.fail(); ++i) {
237 auto tmp = random_internal::read_floating_point<double>(is);
238 if (is.fail()) return is;
239 p.push_back(tmp);
240 }
241 }
242 x.param(param_type(p.begin(), p.end()));
243 return is;
244 }
245
246 ABSL_NAMESPACE_END
247 } // namespace absl
248
249 #endif // ABSL_RANDOM_DISCRETE_DISTRIBUTION_H_
250