MERLIN is the Multi-task Environments for Reinforcement LearnINg generator. This is a fairly straightforward piece of code that generates random problem instances for a multi-task reinforcement learning agent. The main feature it provides is some controls over the structure of the generated problems, such as the “smoothness” of the transition functions of the underlying MDP, the degree to which the individual tasks are correlated/anti-correlated. Some preliminary data from this work was presented at the IEEE Symposium on Adaptive Dynamic Programming and Reinforcement Learning (ADPRL-14) in 2014.
Read More