A reinforcement learning approach for the circle agent of geometry friends

Geometry Friends (GF) is a physics-based platform game, used in one of the AI competitions of the IEEE CIG Conference in 2013 and 2014. The game engages two characters, a circle and a rectangle, in a cooperative challenge involving collecting a set of objects in a 2D platform world. In this work, we propose a novel learning approach to the control of the circle character that circumvents the excessive specialization to the public levels in the competition observed in the other existing solutions for GF. Our approach proposes a method that partitions solving a level of GF into three sub-tasks: solving one platform (SP1), deciding the next platform to solve (SP2) and moving from one platform to another (SP3). We use reinforcement learning to solve SP1 and SP3 and a depth-first search to solve SP2. The quality of the agent implemented was measured against the performance of the winner of the Circle Track of the 2014 GF Game AI Competition, CIBot. Our results show that our agent is able to successfully overcome the over-specialization to the public levels, showing comparatively better performance on the private levels.