int main()
{
using namespace boost::lambda;
static const int num_candidates=60; // divisible by 1,...,candidate::num_traits
static const int N=20000;
typedef strategy<num_candidates> strategy_t;
typedef strategy_t::candidates_t candidates_t;
typedef strategy_t::const_candidate_iterator const_candidate_iterator;
std::vector<boost::shared_ptr<strategy_t> > strategies;
for(int t=1;t<=candidate::num_traits;++t){
strategies.push_back(boost::shared_ptr<strategy_t>(
new simple_strategy<num_candidates>(t)));
}
std::vector<boost::array<int,num_candidates> > histogram(strategies.size());
std::vector<double> acc_score(strategies.size()+1);
for(int n=N;n--;){
typedef boost::array<candidate,num_candidates> candidates_t;
typedef boost::array<const candidate*,num_candidates> candidates_ptr_t;
const candidates_t candidates;
candidates_ptr_t sorted_candidates;
for(std::size_t i=candidates.size();i--;){
sorted_candidates[i]=&candidates[i];
}
std::sort(
sorted_candidates.begin(),sorted_candidates.end(),
bind(&candidate::score,*_1) > bind(&candidate::score,*_2));
acc_score.back()+=sorted_candidates[0]->score;
for(std::size_t t=0;t<strategies.size();++t){
const_candidate_iterator it=strategies[t]->select(candidates);
acc_score[t]+=it->score;
candidates_ptr_t::iterator itp=
std::find_if(
sorted_candidates.begin(),sorted_candidates.end(),
_1==&*it);
std::size_t choice=(std::size_t)(itp-sorted_candidates.begin());
++(histogram[t][choice]);
}
}
for(std::size_t t=0;t<strategies.size();++t){
std::cout<<strategies[t]->name()<<"\t"<<acc_score[t]/N<<std::endl;
for(std::size_t i=0;i<histogram[t].size();++i){
std::cout<<histogram[t][i]<<"\t";
}
std::cout<<std::endl;
}
std::cout<<"optimum_strategy\t"<<acc_score[strategies.size()]/N<<std::endl;
}
void StatisticsBasedScorer::score(const candidates_t& candidates, const diffs_t& diffs,
statscores_t& scores) const
{
if (!m_score_data) {
throw runtime_error("Score data not loaded");
}
// calculate the score for the candidates
if (m_score_data->size() == 0) {
throw runtime_error("Score data is empty");
}
if (candidates.size() == 0) {
throw runtime_error("No candidates supplied");
}
int numCounts = m_score_data->get(0,candidates[0]).size();
vector<int> totals(numCounts);
for (size_t i = 0; i < candidates.size(); ++i) {
ScoreStats stats = m_score_data->get(i,candidates[i]);
if (stats.size() != totals.size()) {
stringstream msg;
msg << "Statistics for (" << "," << candidates[i] << ") have incorrect "
<< "number of fields. Found: " << stats.size() << " Expected: "
<< totals.size();
throw runtime_error(msg.str());
}
for (size_t k = 0; k < totals.size(); ++k) {
totals[k] += stats.get(k);
}
}
scores.push_back(calculateScore(totals));
candidates_t last_candidates(candidates);
// apply each of the diffs, and get new scores
for (size_t i = 0; i < diffs.size(); ++i) {
for (size_t j = 0; j < diffs[i].size(); ++j) {
size_t sid = diffs[i][j].first;
size_t nid = diffs[i][j].second;
size_t last_nid = last_candidates[sid];
for (size_t k = 0; k < totals.size(); ++k) {
int diff = m_score_data->get(sid,nid).get(k)
- m_score_data->get(sid,last_nid).get(k);
totals[k] += diff;
}
last_candidates[sid] = nid;
}
scores.push_back(calculateScore(totals));
}
// Regularisation. This can either be none, or the min or average as described in
// Cer, Jurafsky and Manning at WMT08.
if (m_regularization_type == NONE || m_regularization_window <= 0) {
// no regularisation
return;
}
// window size specifies the +/- in each direction
statscores_t raw_scores(scores); // copy scores
for (size_t i = 0; i < scores.size(); ++i) {
size_t start = 0;
if (i >= m_regularization_window) {
start = i - m_regularization_window;
}
const size_t end = min(scores.size(), i + m_regularization_window + 1);
if (m_regularization_type == AVERAGE) {
scores[i] = score_average(raw_scores,start,end);
} else {
scores[i] = score_min(raw_scores,start,end);
}
}
}