From August, 26th – 29th, 2014, the International Conference "Computational Intelligence in Games" (CIG'2014) will be held in Dortmund. We are happy that we could place a presentation there.
Together with master students Markus Thill, Samineh Bagheri and PostDoc Patrick Koch, I will present some work on Connect-4, the well-known board game. In 2012 we could for the first time construct an agent which learned to play the game nearly perfectly just from self-play. So far, however, learning was rather slow (1.5 millions of training games). Now we could develop with new methods (Online Learning Rate Adaption, Eligibility Traces) a new agent with learning time greatly reduced by a factor of more than 10 (115 000 training games).
The publication is available here ThillCIG2014