A Hierarchical Approach to Multi Tasking in Reinforcement Learning

Abstract

Our aim was to study and evaluate the performance Hierarchical Reinforcement Learning frameworks in multi-tasking domains using active sampling. We evaluated the performance of dmakian implementation of Feudal Network architecture(generate temporally extended sub-policies) on multiple Atari games using OpenAI Gym environment. We integrated multi-tasking algorithms such as Adaptive Active Sampling, Doubling UCB and Doubling DQN for active selection of games during training into the Option Critic architecture using ALE. Future work involves using prioritized experience replay for Doubling DQN and assignment of CPU threads per game based on selection probabilities.