logs4:See if RL framework is working

See if RL framework is working 201804/15 (paused)

• Log 1: what specific output am I working on right now?
  • Before we go with complex RL, we should confirm if RL framework is working.
• Log 2: thinking out loud - e.g. hypotheses about the current problem, what to work on next
  • One of the easiest way is giving higher reward when reply length is longer.
  • by measuring average length by step, we should observe increase of average length.
  • steps
    • enable length reward
    • log average length
    • start RL
• Log 3: record of currently ongoing runs along with a short reminder of what question each run is supposed to answer
  • if avg_len goes up eventually this is good sign.
  • reward also should go up
  • loss should converge
  • if avg_len don't go up, there's something wrong :(
  • Someone taught me reward should be averaged, because it has high variance.
  • Paused here, because we have to fix log issue to see the results.
• Log 4: results of runs (TensorBoard graphs, any other significant observations), separated by type of run (e.g. by environment the agent is being trained in)
  • TODO. This should be logged carefully.
  • TODO model/README + ipnyb
  • models20180414ReinforcementLearningLengthReward