On the coordination system for the dimensionality-reduced inputs of mario

We have proposed neuroevolution method with the dimension-reduced inputs by Manifold Learning for the scene information in Mario AI competition. The merit of using Manifold Learning method is that we can use all the cell in scene for the inputs of Mario AI agents. In this study, we examine two coordination systems: the absolute coordination system and the relative coordination system. In the case of the relative coordination system, the input is the same as the one of simulator provided by the organizer of the competition. The input of the absolute coordination system is generated by the input of the relative coordination system by shifting some rows such that the bottom row is set to be the bottom of ground. Experimental results on the Mario AI show the effectiveness of the absolute coordination system.