16 January. 2021
The availability of huge datasets has enabled the success of deep neural networks trained with supervised learning in problems spanning vision, speech and natural language processing. Reinforcement learning, although it has seen a number of impressive applications (e.g., in games and certain robotic tasks), has yet to be as successful and widespread as supervised learning, due to its online nature. Creating such algorithms that effectively use offline data holds promise in many domains, from autonomous robots to healthcare. Moreover, identifying the hierarchical task decomposition from such data could permit more efficient learning and planning in an online setting. This project will investigate this open research problem using games.
Good programming skills (C++ or Python). Machine learning (reinforcement learning and deep learning) knowledge is a must. Knowledge of hierarchical reinforcement learning, or Unity ML Agents would be considered an advantage.