wiki:BossaExampleThree

Example 3: Calibration jobs and adaptive replication

This example uses a more refined replication policy: instead of doing 2-fold replication, we estimate the error rate of volunteers using calibration tasks, and do enough instance of each job so that the overall probability of error is below a threshold.

We'll track the error rate separately for positive and negative jobs (i.e., images with and without an ellipse). The opaque data structure for users has components:

nneg: # of negative calibration jobs completed
nneg_err: of those, the number of errors
npos: # of positive calibration jobs completed
npos_err: of those, the number of errors

From these we'll derive:

neg_err_rate: error rate for negative cases
pos_err_rate: error rate for positive cases

Our replication policy is:

  • A job is marked Done if either
    • There are N positive instances that match within 20 pixels, and for which the product of pos_err_rate (for the corresponding users) is less than 1e-3, or
    • there are N negative instances and the product of neg_err_rate is less than 1e-3.
  • Else a job is marked Inconclusive if there are 10 finished instances

The job distribution policy is the same as in example 2.

Setup

Create an application named bossa_example3. Create some jobs:

php bossa_example_make_jobs.php --app_name bossa_example3 --dir example

We'll also need to create some calibration jobs:

php bossa_example_make_jobs.php --app_name bossa_example3 --dir example --calibration

This will create 10 calibration jobs based on the images in example/ (recall that these images have corresponding "answer" files.

Callback functions

This example's call back functions are in (html/inc/bossa_example3.inc) The replication policy is implemented in job_finished():

    50  function job_finished($job, $inst, $user) {
    51      $response = null;
    52      if (get_str('submit', true)) {
    53          $response->have_ellipse = 0;
    54      } else {
    55          $response->have_ellipse = 1;
    56          $response->cx = get_int('pic_x');
    57          $response->cy = get_int('pic_y');
    58      }
    59      $inst->set_opaque_data($response);
    60
    61      // if this is a calibration job, update user's opaque data
    62      //
    63      if ($job->calibration) {
    64          $b = $user->bossa;
    65          $info = $job->get_opaque_data();
    66          $answer = $info->answer;
    67          $u = $b->get_opaque_data();
    68          if (!$u) {
    69              $u->npos = 0;
    70              $u->npos_err = 0;
    71              $u->nneg = 0;
    72              $u->nneg_err = 0;
    73          }
    74          if (compatible($response, $answer)) {
    75              if ($answer->have_ellipse) {
    76                  $u->npos++;
    77              } else {
    78                  $u->nneg++;
    79              }
    80          } else {
    81              if ($answer->have_ellipse) {
    82                  $u->npos++;
    83                  $u->npos_err++;
    84              } else {
    85                  $u->nneg++;
    86                  $u->nneg_err++;
    87              }
    88          }
    89          $b->set_opaque_data($u);
    90          return;
    91      }
    92
    93      // now see if job is done
    94      //
    95      $insts = $job->get_finished_instances();
    96      $n = count($insts);
    97
    98      $results = null;
    99      $users = null;
   100      foreach ($insts as $inst) {
   101          $results[] = $inst->get_opaque_data();
   102          $u = $inst->get_user();
   103          $users[] = $u->bossa->get_opaque_data();
   104      }
   105
   106      // see if there's a negative consensus
   107      //
   108      $prob = 1;
   109      for ($i=0; $i<$n; $i++) {
   110          $r = $results[$i];
   111          if ($r1->have_ellipse) continue;
   112          $u = $users[$i];
   113          $prob *= $u->neg_err_rate;
   114      }
   115      if ($prob < PROB_LIMIT) {
   116          $job->set_state(BOSSA_JOB_DONE);
   117          return;
   118      }
   119
   120      // see if there's a positive consensus
   121      //
   122      for ($i=0; $i<$n; $i++) {
   123          $r1 = $results[$i];
   124          $u = $users[$i];
   125          $prob = $u->pos_error_rate;
   126          for ($j=0; $j<$n; $j++) {
   127              if ($j == $i) continue;
   128              $r2 = $results[$j];
   129              if (compatible($r1, $r2)) {
   130                  $u2 = $users[$j];
   131                  $prob *= $u2->pos_err_rate;
   132              }
   133          }
   134          if ($prob < PROB_LIMIT) {
   135              $job->set_state(BOSSA_JOB_DONE);
   136              return;
   137          }
   138      }
   139
   140      // see if there are too many instances without a consensus
   141      //
   142      if ($n >= 10) {
   143          $job->set_state(BOSSA_JOB_INCONCLUSIVE);
   144          return;
   145      }
   146
   147      // still looking for consensus - get another instance
   148      //
   149      $job->set_priority(2);
   150
   151  }

We also supply a callback function to show a user's opaque data on administrative web pages:

function user_summary($user) {
    $b = $user->bossa;
    $info = $b->get_info();
    if ($info) {
        if ($info->npos) {
            $pos_err = $info->npos_err/$info->npos;
        } else {
            $pos_err = "---";
        }
        if ($info->nneg) {
            $neg_err = $info->nneg_err/$info->nneg;
        } else {
            $neg_err = "---";
        }
        return "error rate: positive $pos_err ($info->npos_err/$info->npos),
             negative $neg_err ($info->nneg_err/$info->nneg)
        ";
    } else {
        return "No data";
    }
}
Last modified 9 years ago Last modified on 08/19/08 16:06:09