Success stories of deep reinforcement learning (RL) from high dimensional inputs such as pixels or large spatial layouts include achieving superhuman performance on Atari games [30, 37, 1], grandmaster levelinStarcraft II[50]andgrasping adiverse setofobjects with impressivesuccess rates and generalization with robots in the real world [21].

Offline safe reinforcement learning (RL) aims to find an optimal policy using a pre-collected dataset when data collection is impractical or risky.

Neural Information Processing Systems http://nips.cc/

See Table 1for the average running time per problem instance. Note that the implementation of Z3 and OR-tools22 are in C++, while NeuRewriter and RL baselines are in Python. Still, we can observethat our approach achieves a23 better balance between the time-efficiency and the result quality. For expression simplification and job scheduling,24 NeuRewriter is even more time-efficient than Z3 and OR-tools. The region-pickerπω is parameterized by aQ-function and is similar in spirit to soft-Q learning [2].