Using visual recognition software while processing video clips of people playing Connect 4, Gomoku, Pawns and Breakthrough — including games ending with wins, ties or those left unfinished — the system would recognise the board, the pieces and the different moves that lead to each outcome. A unique formula then enabled the system to examine all viable moves when playing and, using data gathered from all possible outcomes, calculate the most appropriate move. Łukasz Kaiser of Paris Diderot University developed the learning algorithms after noticing a glaring gap in our knowledge — object recognition machine learning experiments are fairly popular, however studies into high-concept computer learning are less common, despite having plenty of future uses in the creation of autonomous robots. Kaiser chose to use games as a primary learning tool because they are, “a natural model of many real-world interaction scenarios, making the results significant in a broader context”.
By forgoing an initial singular formula and using relational structures that recognise the rows, columns and diagonals of a boardgame, and making use of several different logic systems — including pure first-order, existential and guarded — a tailored formula could be devised from the data gathered from each logic. An added General Game Playing program helped the system learn how to play tactically, learn legal moves and, ultimately, win.
“This combination allowed it to generate very short and intuitive formulas in the experiments we performed, and there is strong theoretical evidence that it will generalise to other problems,” concluded the paper. Kaiser plans to adapt the system to solve problems that require “hierarchical, structured learning or a form of probabilistic formulas”, all of which will come in handy in developing autonomous, intelligent robots.
