Memory based Multi-tasking A3C Agent

Abstract

Reinforcement learning agents have been found to be successful in various tasks and have even been transferred to and used in real-world domains. However, there is still a long way to go before they come close to humans when it comes to multi-tasking and transfer learning over similar and vastly different tasks. The idea for our project emerged from the belief that memory storage in cells is one advantage humans possess over machines that results in a significant difference in performance level, especially in data efficiency and long term dependencies between rewards and actions in a Reinforcement learning setup. In this project, we have done an extensive survey on existing methods for multi-tasking and incorporating memory in RL agents, and have developed and deployed a model that has memory and can learn to do multiple tasks. We have demonstrated the performance of our model on single as well as multiple(multi-tasking setup) Atari games, but in theory, the model can be extended to any reinforcement learning problem.