Optimization of Angry Birds AI controllers with distributed computing

The one of important issues in artificial intelligence (AI) research is the development of AI for games because of its difficulty. To promote the research on video games AI, there have been several game AI competitions. However, some games with physics engine (geometry friends or Angry Birds) have no support on the prediction of future events using simulation. It makes much difficult to build AI for the games with physics. As a result, AI creator should spend much time to optimize the parameters of their program by trial and errors. In this paper, we report our approach to build AI for Angry Birds (Plan A+, 3rd rank in 2014 Angry Birds AI competition and the first entry achieved 1 million points in benchmarking test). In our controller, we adopt multiple strategies to increase generalization ability and hybrid optimization techniques (greedy search from human's manually tuned parameters) with parallel machines.