Dynamic Network Architectures for Deep Q-Learning: Modelling Neurogenesis in Artificial Intelligence

Abstract

Artificial neural networks have become popular within a range of machine learning fields for their ability to solve complex problems, with one of the uses as function approximators in Q-learning. These networks generally have static architectures, which is a problem in the regard of artificial general intelligence, since no single specific architecture is optimal for all problems. In this thesis, we implement and evaluate a proof of concept for a novel approach of a dynamic network architecture, resulting in a model that can be seen as a combination of compressed classical table-based Q-learning and artificial neural networks. The model presented performs true tabula rasa deep Q-learning, starting with an empty network that is gradually extended with nodes when experiencing “surprising” events, and is capable of generalization by abstracting important features from noisy input. Finally, we show that the model can learn from delayed rewards in simple environments and compare it with the well-established DQN algorithm.