Learning Similarity Metrics Between Simulation and the Real World (2020)

Using knowledge gathered from simulated environments in the real world can be seen as a transfer learning approach where simulated environments are the source tasks and the real world is the target task. The accuracy of the transfer depends on how similar the source simulations and the real world are. From the Reinforcement Learning point of view, that transfer can be understood as how useful the policies learned in the simulations are for solving the problems of the real world. Probabilistic Policy Reuse (PPR) raised as a transfer learning method that was able to accurately reuse policies learned in source tasks (in this case, simulation) in a target task (the real wold). In addition, one of the main characteristics of PPR is that it was able to determine a distance metric between the different tasks. In this project, we plan to extend PPR concepts to find accurate similarity metrics between simulation and the real world.Read more about the project here.

PI: [Principal Investigator]

Core team: PLG and JP Morgan

Funding: This project is granted by JP Morgan