ETD

• multi-task RL
• policy transfer
• reinforcement learning

This thesis proposes a new Multi Task Reinforcement Learning approach to improve the learning of a new task by using previously acquired knowledge (source policies). This method leverage a common latent space for frames and task descriptions, using an Autoencoder and BERT transformer, to capture the similarity between source and target tasks via cosine similarity. The closest source policy is selected to guide the learning process of the target task. The experiments were carried out on two different game environments: Atari and Highway-env.