If you think you’re good at board games, you just met your match. Because a computer scientist has developed machine learning software that can watch your playing for just two minutes — then comprehensively thrash you.
Łukasz Kaiser, a computer scientist from Paris Diderot University, has developed a machine learning algorithm [PDF] that uses visual recognition software to recognise the board and pieces of different games. During the processes, it crunches the data corresponding to different moves in order to work out how they affect the outcome.
Supplied with a few rules of the game, it can then analyse all viable moves when playing and, using data gathered from watching a player for two minutes, calculate the most appropriate move. So far, the software can wipe the floor with people playing Connect 4, Gomoku, Pawns and Breakthrough.
This isn’t some IBM-style supercomputer system, either: Kaiser in fact runs the software on a laptop with just 4GB RAM and a single processor core. The reason that’s possible is because he’s junked the theory used by previous game solvers that relied on inductive logic, instead using relational structures that recognise the rows, columns and diagonals of a board game, then makes use of several different logic systems.
Of course, Connect 4 doesn’t sound much of a challenge — which is why he plans to investigate games that require “hierarchical, structured learning or a form of probabilistic formulas” in the future. Sadly, he won’t let on which games he’s thinking of; let’s just hope he stays the hell away from Global Thermonuclear War.. [Łukasz Kaiser via Wired]