Abstract by Joshua Greaves
3M Holodeck: A High Fidelity Simulator For Deep Reinforcement Learning
Despite recent advances in reinforcement learning and deep neural networks, reinforcement learning algorithms still require a large amount of time to perform well in relatively simple domains. However, training in faster than real time simulation can dramatically reduce the time needed for these algorithms to learn useful policies, especially when these simulations can run in parallel. Furthermore, there are an increasing number of tasks of real-world complexity that we would like to be able to learn algorithms for. To this end, we introduce Holodeck, a high fidelity faster than real time simulator. Built on top of Unreal Engine 4, Holodeck enables the easy incorporation of the latest computer graphics and fast inter-process communication methods to allow interaction with the simulated environments through Python.