Abstract
Researchers have made a lot of progress in combining the advances in Deep Learning and the generalization and applicability of Reinforcement learning to the sequential decision-making process and introduce Deep Reinforcement Learning, which allows using Reinforcement Learning in high dimensional input space environments. Deep Reinforcement Learning achieved notable results in Atari-Games, continuous control tasks such as Robotics. In this project we benchmark the performance of three different deep reinforcement Learning techniques Deep Deterministic Policy Gradient “DDPG” [1], Deep Deterministic Policy Gradient with Hindsight Experience Replay “DDPG + HER” [2] and state-of-art policy gradient method Proximal Policy optimization “PPO” [3], on multi-goal continuous control environments Fetch task and HandManipulate tasks, we benchmarked the three algorithms on six different environments using sparse and dense reward settings. Deep Deterministic Policy Gradient with Hindsight Experience Replay [2] achieves the best success-rate overall the environments when applied with sparse rewards, while both Proximal Policy Optimization [3] and Deep Deterministic Policy Gradient [1] were able to converge only on FetchReach environment.
Benjamin Rosman
Lab Director
I am a Professor in the School of Computer Science and Applied Mathematics at the University of the Witwatersrand in Johannesburg. I work in robotics, artificial intelligence, decision theory and machine learning.