Semi Unsupervised Meta Reinforcement Learning

Abstract

The field of Reinforcement Learning (RL) has been developing extremely fast in the last few years. Its combination with Neural Networks has transformed this area of Machine Learning into a powerful tool that can solve a variety of problems with efficiency and many times without the help of humans. In this thesis, we created an end-to-end learning algorithm (SUMRL) by combining RL and Meta Learning techniques for fast and efficient adaptation to new and unknown tasks. In order to free our agent as much as possible from human supervision, we implemented a technique of Unsupervised Learning (with the help of the Information Theory field) and motivated our agent to create unsupervised skills. These skills were later fed to Meta Learning algorithms as tasks for our agent to meta-train on. We evaluated our algorithm by comparing the meta-testing outcome when the agent was meta-trained with our pretrained skills/tasks and with random tasks. Our experiments showed that although our agent has a good overall performance in both cases, pretrained tasks were not able to clearly return better results than the random tasks at the process of meta-testing.