Nice overview of reinforcement learning and the state of its development right now.
Agarwal compares the availability of experimentation platforms for reinforcement learning to the impact large labeled data sets like ImageNet had on supervised learning. “The way we make a lot of progress in supervised learning was that we started accumulating large data sets and repositories and once we had those, we could try algorithms out on them reliably and iterate those algorithms.” A static data set isn’t useful for evaluating more general reinforcement learning; “two different agents will take two different trajectories through an environment.”
Instead, researchers need a large, diverse set of environments that’s also standardized so everyone in the field works against them. “Flexible, diverse platforms can serve the same function as a repository for reinforcement learning tasks where we can evaluate and iterate on ideas coming out of research much faster than was possible in the past, when we had to restrict the algorithms to simple evaluation problems because more complex ones weren’t available. Now we can take ideas to the platforms and see whether or not they do a good job,” Agarwal said.
Originally shared by Radhika Jadcherla
A good read…
Share your comments