Timezone: »

Efficient Offline Policy Optimization with a Learned Model
Zichen Liu · Siyi Li · Wee Sun Lee · Shuicheng Yan · Zhongwen Xu
Event URL: https://openreview.net/forum?id=nyo7Y-CJRGf »

MuZero Unplugged presents a promising approach for offline policy learning from logged data. It conducts Monte-Carlo Tree Search (MCTS) with a learned model and leverages Reanalyze algorithm to learn purely from offline data. For good performance, MCTS requires accurate learned models and a large number of simulations, thus costing huge computing time. This paper investigates a few hypotheses where MuZero Unplugged may not work well under the offline RL settings, including 1) learning with limited data coverage; 2) learning from offline data of stochastic environments; 3) improperly parameterized models given the offline data; 4) with a low compute budget. We propose to use a regularized one-step look-ahead approach to tackle the above issues. Instead of planning with the expensive MCTS, we use the learned model to construct an advantage estimation based on a one-step rollout. Policy improvements are towards the direction that maximizes the estimated advantage with regularization of the dataset. We conduct extensive empirical studies with BSuite environments to verify the hypotheses and then run our algorithm on the RL Unplugged Atari benchmark. Experimental results show that our proposed approach achieves stable performance even with an inaccurate learned model. On the large-scale Atari benchmark, the proposed method outperforms MuZero Unplugged by 43%. Most significantly, it uses only 5.6% wall-clock time (i.e., 1 hour) compared to MuZero Unplugged (i.e., 17.8 hours) to achieve a 150% IQM normalized score with the same hardware and software stacks.

Author Information

Zichen Liu (national university of singaore, National University of Singapore)
Siyi Li (Sea AI Lab)
Wee Sun Lee (National University of Singapore)

Wee Sun Lee is a professor in the Department of Computer Science, National University of Singapore. He obtained his B.Eng from the University of Queensland in 1992 and his Ph.D. from the Australian National University in 1996. He has been a research fellow at the Australian Defence Force Academy, a fellow of the Singapore-MIT Alliance, and a visiting scientist at MIT. His research interests include machine learning, planning under uncertainty, and approximate inference. His works have won the Test of Time Award at Robotics: Science and Systems (RSS) 2021, the RoboCup Best Paper Award at International Conference on Intelligent Robots and Systems (IROS) 2015, the Google Best Student Paper Award, Uncertainty in AI (UAI) 2014 (as faculty co-author), as well as several competitions and challenges. He has been an area chair for machine learning and AI conferences such as the Neural Information Processing Systems (NeurIPS), the International Conference on Machine Learning (ICML), the AAAI Conference on Artificial Intelligence (AAAI), and the International Joint Conference on Artificial Intelligence (IJCAI). He was a program, conference and journal track co-chair for the Asian Conference on Machine Learning (ACML), and he is currently the co-chair of the steering committee of ACML.

Shuicheng Yan (Sea AI Lab)
Zhongwen Xu (Sea AI Lab)

More from the Same Authors