The success of multi-task deep reinforcement learning has been limited so far and the community is currently lacking explanations. We suspect that the main reasons are instabilities between gradients that are coming from different tasks.
We support these claims with results from representation learning, a powerful tool to discover properties of learning algorithms. In particular, t-SNE has been demonstrated to be useful for visualizing the learned representation of Deep Reinforcement Learning (DRL) agents, showing an ability to interpret learned policies.
Below are two different t-sne maps that are related to multi-task deep reinforcement learning.
Rusu et al. (2015) (Policy distillation)- claim that multi-task DRL is hard. Their method is to first train teacher networks and then compress them into a single net
