Achievement

Recent research in reinforcement learning leverages results from the “continuous-armed bandit” literature to create algorithms for continuous state and action spaces. Trainee Ariel Weinstein working with faculty advisor Littman examined the empirical properties of algorithms that plan in an open-loop manner, concerned only with a sequence of actions and not state information. Therefore the algorithms are completely agnostic to state and can function in domains with discrete states, continuous states or a mixture of both. Weinstein demonstrated the effectiveness of this planning method when coupled with exploration and model learning and show that the approach is very competitive with other continuous-action reinforcement learners. (“Bandit-Based Planning and Learning in Continuous-Action Markov Decision Processes”, To appear in International Conference on Automated Planning and Scheduling 2012).