Abstract
Geometry Friends (GF) is a physics-based platform game, which was part of the Artificial Intelligence (AI) competitions of the IEEE CIG Conference in 2013 and 2014. On GF there are two different characters, a circle and a rectangle, whose goal is to catch all the diamond-shaped collectibles available on each level of the game. In this work, a novel approach to the GF problem for the circle agent is proposed. This approach is based on learning algorithms, is character-agnostic and circumvents the excessive specialization to the public levels observed in the agents submitted to the 2014 competition. The solution uses a Divide-and-Conquer strategy that partitions the problem of solving a GF level into a series of three sub-problems: solving one platform (SP1), deciding the next platform (SP2) and moving from one platform to another (SP3). This method uses reinforcement learning to solve SP1 and SP3 and a depth-first search to solve SP2. To measure the quality of the developed agent, its results on the levels of the 2014 Competition are measured against the performance of that competition contestants, CIBot and KUAS-IS Lab The results show that despite having a worse performance overall, the agent successfully avoided becoming over-specialized to a specific sub-set of levels.