Reinforcement Learning Game – RLGame


Scientists from Hellenic Open University and Computer Technology Institute and Press “Diophantus” have used the HellasGrid Infrastructure and the EGI Grid infrastructure in order to solve problems coming from the area of Machine Learning and Algorithms in Games.

When using machine learning to learn how to play a game, vast amounts of experiments may be required to adequately explore the space of alternative tactics and strategies. To cover this need, grid was a technology which was exploited by the researchers.

The scientists have identified a curious phenomenon in game playing, which they term “the pendulum effect”. Therein, they use a tutor to provide playing advice to one player (A) of a zero-sum two-player game. It seems that the player (B) who does not receive tutoring is also able to enormously benefit just by being forced to play against a better player (A, as advised by the tutor). This is reinforced when the tutor abstains; player B seems to be able to win more often as compared to when the tutor is present directing A to more wins.

By using the HellasGrid and EGI infrastructure, the scientists conducted a range of experiments across hundreds of game configurations and across several learning schemes; some of them employed game tree look-ahead search which is pretty expensive (but not yet optimized for grid computing). These experiments may have consumed several tens of thousands of CPU hours on the grid; such availability of resources simply does not exist at stand-alone venues.

Future plans include the continuation of this line of research to verify this type of learning behavior, across more configurations and actions to verify it across other games as well. The scientists are also currently looking into workflows (hopefully, grid-aware workflow systems) to better organize data collection and analysis.

Contacts

  • Dimitris Kalles, Hellenic Open University, kalles (at) eap.gr
  • Panagiotis Kanellopoulos, Computer Technology Institute and Press “Diophantus”, kanellop (at) ceid.upatras.gr
References
  1. D. Kalles, and I. Fykouras. “Examples as Interaction: On Humans Teaching a Computer to Play a Game”, International Journal on Artificial Intelligence Tools, 2010.
  2. D. Kalles and P. Kanellopoulos. “A Minimax Tutor for Learning to Play a Board Game”, Workshop on Artificial Intelligence in Games, a workshop of the 18th European Conference on Artificial Intelligence, 2008.
  3. D. Kalles and P. Kanellopoulos. “A Pendulum Effect in Co-evolutionary Learning in Games”, European Workshop in Reinforcement Learning, 2011.