Addressing State Representation in Deep Reinforcement Learning: a critical analysis of state-of-the-art” methods

Deep Reinforcement Learning models use a Deep Neural Network to compute the Q-function, avoiding some computational and memory issues related to the Q-table in classic Reinforcement Learning. However, DeepRL models suffer from high sample complexity; to deal with this problem, it has been recently proposed to exploit the Markov Decision Process underlying DeepRL models, allowing to use Graph Representation Learning approaches to obtain efficient state representations to feed the Deep Neural Network with. Following this idea, we focus on random walk based GRL: the main methods in this category are analyzed and compared, and the most suitable one is used for graph representation in a DeepRL model, whose results are then discussed and compared to the ones of a DeepRL model without GRL techniques.