Alvin Wan Rose Hills
Evolution Strategies as Derivative Free Alternative to Deep Reinforcement Learning
Deep Reinforcement Learning (DRL) has seen state-of-the-art results with Atari games, spoken dialogue systems, and a differentiable neural computer. However, excessive amounts of computer power are required to attain such results. My work concerns potential simplifications of DRL so that more advanced tasks are feasible, leveraging alternative evolutionary strategies with deep function approximators and evolutionary strategies with linear function approximators.