1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
|
#include "oaa.h"
#include "vw.h"
#include "cb.h"
#include "rand48.h"
using namespace LEARNER;
namespace CBIFY {
struct cbify {
size_t k;
size_t tau;
float epsilon;
CB::label cb_label;
};
void do_uniform(cbify* data, example* ec)
{
//Draw an action
uint32_t action = (uint32_t)ceil(frand48() * data->k);
ec->final_prediction = (float)action;
}
void do_loss(example* ec)
{
OAA::mc_label* ld = (OAA::mc_label*)ec->ld;//New loss
if (ld->label != ec->final_prediction)
ec->loss = 1.;
else
ec->loss = 0.;
}
template <bool is_learn>
void predict_or_learn_first(cbify* data, learner& base, example* ec)
{//Explore tau times, then act according to optimal.
OAA::mc_label* ld = (OAA::mc_label*)ec->ld;
//Use CB to find current prediction for remaining rounds.
if (data->tau > 0)
{
do_uniform(data, ec);
do_loss(ec);
data->tau--;
cout << "tau--" << endl;
uint32_t action = (uint32_t)ec->final_prediction;
CB::cb_class l = {ec->loss, action, 1.f / data->k};
data->cb_label.costs.erase();
data->cb_label.costs.push_back(l);
ec->ld = &(data->cb_label);
if (is_learn)
base.learn(ec);
else
base.predict(ec);
ec->final_prediction = (float)action;
ec->loss = l.cost;
}
else
{
data->cb_label.costs.erase();
ec->ld = &(data->cb_label);
if (is_learn)
base.learn(ec);
else
base.predict(ec);
do_loss(ec);
}
ec->ld = ld;
}
template <bool is_learn>
void predict_or_learn_greedy(cbify* data, learner& base, example* ec)
{//Explore uniform random an epsilon fraction of the time.
OAA::mc_label* ld = (OAA::mc_label*)ec->ld;
data->cb_label.costs.erase();
ec->ld = &(data->cb_label);
base.predict(ec);
do_loss(ec);
uint32_t action = (uint32_t)ec->final_prediction;
float base_prob = data->epsilon / data->k;
if (frand48() < 1. - data->epsilon)
{
CB::cb_class l = {ec->loss, action, 1.f - data->epsilon + base_prob};
data->cb_label.costs.push_back(l);
}
else
{
do_uniform(data, ec);
do_loss(ec);
action = (uint32_t)ec->final_prediction;
CB::cb_class l = {ec->loss, (uint32_t)ec->final_prediction, base_prob};
data->cb_label.costs.push_back(l);
}
if (is_learn)
base.learn(ec);
ec->final_prediction = (float)action;
ec->loss = data->cb_label.costs[0].cost;
ec->ld = ld;
}
void learn_bagging(void* d, learner& base, example* ec)
{//Randomize over predictions from a base set of predictors
// cbify* data = (cbify*)d;
//Use CB to find current predictions.
}
void learn_cover(void* d, learner& base, example* ec)
{//Randomize over predictions from a base set of predictors
//cbify* data = (cbify*)d;
//Use cost sensitive oracle to cover actions to form distribution.
}
void finish_example(vw& all, cbify*, example* ec)
{
OAA::output_example(all, ec);
VW::finish_example(all, ec);
}
learner* setup(vw& all, std::vector<std::string>&opts, po::variables_map& vm, po::variables_map& vm_file)
{//parse and set arguments
cbify* data = (cbify*)calloc(1, sizeof(cbify));
data->epsilon = 0.05f;
data->tau = 1000;
po::options_description desc("CBIFY options");
desc.add_options()
("first", po::value<size_t>(), "tau-first exploration")
("greedy",po::value<float>() ,"epsilon-greedy exploration");
po::parsed_options parsed = po::command_line_parser(opts).
style(po::command_line_style::default_style ^ po::command_line_style::allow_guessing).
options(desc).allow_unregistered().run();
opts = po::collect_unrecognized(parsed.options, po::include_positional);
po::store(parsed, vm);
po::notify(vm);
po::parsed_options parsed_file = po::command_line_parser(all.options_from_file_argc,all.options_from_file_argv).
style(po::command_line_style::default_style ^ po::command_line_style::allow_guessing).
options(desc).allow_unregistered().run();
po::store(parsed_file, vm_file);
po::notify(vm_file);
if( vm_file.count("cbify") ) {
data->k = (uint32_t)vm_file["cbify"].as<size_t>();
if( vm.count("cbify") && (uint32_t)vm["cbify"].as<size_t>() != data->k )
std::cerr << "warning: you specified a different number of actions through --cbify than the one loaded from predictor. Pursuing with loaded value of: " << data->k << endl;
}
else {
data->k = (uint32_t)vm["cbify"].as<size_t>();
//appends nb_actions to options_from_file so it is saved to regressor later
std::stringstream ss;
ss << " --cbify " << data->k;
all.options_from_file.append(ss.str());
}
all.p->lp = OAA::mc_label_parser;
learner* l = new learner(data, all.l, 1);
if (vm.count("first") )
{
data->tau = (uint32_t)vm["first"].as<size_t>();
l->set_learn<cbify, predict_or_learn_first<true> >();
l->set_predict<cbify, predict_or_learn_first<false> >();
}
else
{
if ( vm.count("greedy") )
data->epsilon = vm["greedy"].as<float>();
l->set_learn<cbify, predict_or_learn_greedy<true> >();
l->set_predict<cbify, predict_or_learn_greedy<false> >();
}
l->set_finish_example<cbify,finish_example>();
cout << "epsilon = " << data->epsilon << endl;
return l;
}
}
|
