Artificial intelligence is mastering more and more complex games
The advances of artificial intelligence become obvious with its ability to compete in more and more complex games: In 1996 IBM‘s Deep Blue reached super-human ability in chess, in 2007 scientists created an unbeatebly checkers AI and just recently AlphaGo bea the Grandmaster in Go, a game that is unlikely more complex than chess and checkers. However these games still are relatively simple compared to many real world challenges. Thus the next challenge would be to develop an AI that can compete in even more complex environments.
The Freeciv Challenge
Freeciv is a strategic game that comes very close to many real world problems. It has much more possible game states than chess or Go and there exists incomplete information, which means that the player cannot observe the progress of their enemies. Furthermore, randomness and the balancing of short term and long term goals make Freeciv even for humans a very challenging task. We are taking on the challenge of mastering the Game of Freeciv with our artificial intelligence HIRO.
In Freeciv the player needs to build up a strong civilization and conquer the rest of the world or dominate it with technology. It is round based, so the player needs to make several hundred of choices in every round: set reserach goals, build and develop cities, plan military operations, etc.
The common approach of many AI systems is to take into account all possible combinations of a game and evaluate their „desirability“. However, this is not possible if the number of possible states is immensely large and if a game state does not provide all necessary information for the next decisions. That‘s why HIRO is building on a knowledge base that incorporates the knowledge of the best Freeciv players in the world. By using the knowledge about best-practices, HIRO can ignore many possible actions that are not relevant to win the game and pick only the best actions necessary to win.
Why HIRO has good chances to win
Unlike other experts systems, HIRO can reuse its knowledge and can be easily trained by many different people at the same time. HIRO learns from language that can be shared among developers and thus makes learning from humans very efficient. In addition HIRO can evaluate its actions and learn from experience which strategies have worked best. These properties give HIRO the capabilities to excel in very complex situations such as IT system management and also in Freeciv.