General Game Playing

My photographThis research project aims towards optimization of general game controllers or more generally solving problems of reinforcement learning. Currently we experiment with games on Atari 2600 platform. The goal is to develop a method able to fully-automatically generate a general controler for a chosen set of games. The contoller takes unpreprocessed game screen images as its input. The output corresponds with joystick movement. We apply state-of-the-art approaches to solve the problem: deep networks (convolutional neural networks), indirect encodings of neural networks, evolutionary algorithms and advanced reinforcement learning techniques. The outcomes of this project might form a stepping stone towards general AI.

See the Atari-ANN blog: