Attendees

• Joe
• Chi Hui

Status

• Bug identified and squashed
  • Our updated "double threshold" single objective policy performs very similarly (according to g1) to the "queue" policy in some cases
  • In other cases, the "queue" policy performs BETTER than the single objective policy (according to g1)
    • This could mean that our new reward definition induces speed behavior similar to that of the queue policy
      • Though the queue policy always performs MUCH BETTER than the double threshold policy according to g2
    • We may have just needed to train the double threshold policies longer (though I doubt it because it seemed like we obtained convergence)
• Reran batch offline learning experiment with "queue" policy and "threshold 1.0/13.89" policy
  • See Experiment 21.1

Next Steps

• Take out square root part of reward definition and regenerate table from experiment 20
  • We need the queue policy to perform much worse than the excess speed policy