Functional Learning-Modules for Atari Ball Games

Andrew Melnik, Sascha Fleer, Malte Schilling, Helge Ritter

Abstract: We present a cognitive architecture capable of end-to-end learning across different but conceptually similar Atari games using the same learning algorithms, network architecture, and hyperparameters. The cognitive architecture uses object-based representations to generalize across different games and consists of a pipeline of functional modules. Each module is allocated for a specific functionality, but exact policies and input-output relationships are formed by reinforcement learning and supervised learning algorithms. The convergence rate of the cognitive architecture is considerably faster than deep reinforcement learning architectures that rely on pixel-based information and follow a purely end-to-end approach. Moreover, the modules of the cognitive architecture can be directly reused in different game environments without retraining. Our approach is inspired by the modular structure of biological brains, where functional modules evolved, but still can be shaped through learning from new experiences.