Learning to Play Pong using Policy Gradient Learning

Abstract

Activities in reinforcement learning (RL) revolve around learning the Markov decision process (MDP) model, in particular, the following parameters: state values, V; state-action values, Q; and policy, pi. These parameters are commonly implemented as an array. Scaling up the problem means scaling up the size of the array and this will quickly lead to a computational bottleneck. To get around this, the RL problem is commonly formulated to learn a specific task using hand-crafted input features to curb the size of the array. In this report, we discuss an alternative end-to-end Deep Reinforcement Learning (DRL) approach where the DRL attempts to learn general task representations which in our context refers to learning to play the Pong game from a sequence of screen snapshots without game-specific hand-crafted features. We apply artificial neural networks (ANN) to approximate a policy of the RL model. The policy network, via Policy Gradients (PG) method, learns to play the Pong game from a sequence of frames without any extra semantics apart from the pixel information and the score. In contrast to the traditional tabular RL approach where the contents in the array have clear interpretations such as V or Q, the interpretation of knowledge content from the weights of the policy network is more illusive. In this work, we experiment with various Deep ANN architectures i.e., Feed forward ANN (FFNN), Convolution ANN (CNN) and Asynchronous Advantage Actor-Critic (A3C). We also examine the activation of hidden nodes and the weights between the input and the hidden layers, before and after the DRL has successfully learnt to play the Pong game. Insights into the internal learning mechanisms and future research directions are then discussed.