Other-Centered Ethics and Harsanyi's Aggregation Theorem

Abstract

This article argues that the progress seen in Reinforcements Learning (RL) field is mostly due to compute, data, and infrastructure rather than algorithmic breakthroughs. It signifies that RL follows a procedure similar to supervised learning, except that the correct labels are replaced with the sampled actions generated by the policy network. The article delves into the fundamental concepts of RL, such as policy networks, gradients, and the credit assignment problem. It presents a walkthrough of training a policy network to play Pong using policy gradients and discusses how advantages can be used to modulate the loss function, enabling learning from delayed rewards. The article also highlights cases where policy gradients fall short compared to human intelligence and concludes with reflections on the broader implications of RL for advancing AI and its practical applications in robotics. – AI-generated abstract.