Reminiscence, Starcraft Winning Prediction 🍊

leader

It’ was the first time to participate so-called a competition in data-analysis field ever, but it was exciting experience and We really were enjoying the time, every moments talking about own thouhgt to one and another. since I’d been motivated while we’re in a small study group for data analysis for a couple of time.



In fact,’ it looks like bumping into every wall and rock bottom, sometimes, We had to try-and-frustration back and forth even over the midnight.



But,We passed the qualifing baseline close call

  • 66.51% - prediction rate, it was ranked 45 out of 669 teams.
  • It was a fresh good start at the beginning but not prolonged any longer.
  • We were overwhelmed by so-called GOD-baseline. (GOD’s cutoff line)
  • No progress couldn’t be made a little bit despite of long-time efforts.



When,We even got in Top-10s with 72.47%, We were thrilled

  • We were curious wheather a few effective powerful features will kick in, or..
  • We don’t know but the features as many as possible would be better.
  • We picked modest one somewhere in between the two.
  • some of strong guesses, plus a many whatever it seems to be good one
  • consequently, we decided to use finally picked ( ) features.



But,the hidden experts always start to move at the last moments

  • We were already out of our submissions, no motivation any more at 72.53%.
  • while We somehow were satisfied a small achivement at the first try,
  • the ranks were disturbing at last day, even until the very last moment.
  • We were pushed out of Top-10%, but keep staying, finally remarked at 10th.



Finalexamining, fully assessmented when the day closed by test-set of 100%

  • the last twist happened in final assessment when the day was closed.
  • during the competition-days, draft assessments were made with 30% test-Data
  • final assessment was fully examined score at 100% test data-set.
  • ranking was disturbed drastically anywhere in leader board.
  • There were 3 or 4 top-10 members were disappered losing their positions
  • a certain higher ranker fell off even under a quialifing cut-off.
  • but we could stay strong on the 10th, that was a achivement we made.



Thesubmissions are limmited to three in a day

  • Way-01 is a case of LGBM, and 02 is XGBM results.
  • We compare the results of two models and parameters.
  • The results were founded __LGBM was a bit more effective__ on this.
  • The iteration was not an essential, sometimes, even be worse.



Sometimes,there’s a UREKA moment, gets a light

  • there seems to be timespans when the player were weak or strong.
  • we evaluated a winning rate depending on a tendancy (stiffness of graph)
  • sometimes, we found out elaborates were all in vain due to wrong hyphothesis.



Thebiggest frustration were computing resources

  • google colaboratory is great, but it isn’t what could deal with bigger data.
  • but plan for price is different. it gives power or money, the reason why we have to pay for.
  • but the power, looked so absolute were revealed cranky. lightening and snail, heaven or hell.



16cores, 32GB RAM server was provide from team member, yeah

  • finally, we met a messanger, quit cool workstation pc was offered.
  • but always worried about shortage of memory, keep eyes on resource monitor.
  • team owned server gave modest resource power, but reliable and steady.



Actually,it wasn’t that powerful, BUT steady perfomance

  • With jupyter server, it was quit neat and comvenient.
  • it was not that google-powerful, but shows steady performance.





Written on ... April 30, 2020