American Go E-Journal

Your Move/Readers Write: An IBM CTO Responds to Scottish Neural Network

Saturday December 20, 2014

“It’s interesting to read about the work of University of Edinburgh to use machine learning to improve the level of playing in computers, (Scottish Neural Network Takes Computer Go to Next Level 12/16/2014 EJ)” writes Nin Lei, Distinguished Engineer and CTO, Analytics and Big Data, STG IBM Systems and Technology Group. “However, the title in their article creates an impression that their research is creating a program that can beat the best human players. If their probability of guessing their next move is only 44%, then their chance of guessing it wrong is 56%.  In a sequence of 10 moves, the chance of getting the complete sequence correctly is 0.44 ** 10, which is a very small number.” Noting that checkers “has been solved via machine learning,” Lei says that “it appears it is promising for go as well.” But because machine learning predicates that there is a pattern in the underlying data set, Lei warns that “it could be so complex that machine learning can only attain a certain level of accuracy.  It seems to me a program needs to have very high level of accuracy before it can play a good game at strong human level.” Lei also says that “Since machine learning is based on pattern recognition, I wonder if a professional can trick the program by using moves that may not be optimally locally but will create patterns that the program has not seen before. I applaud the work they are doing,” Lei concludes. “It is innovative by using a different approach than the existing strong computer programs.  It will be interesting to find out if someday they can come up with an algorithm that can improve the accuracy significantly.”
12/22: the chance of getting the complete sequence correctly has been corrected to 0.44 ** 10 (from 0.56).