Synthesizing Programmatic Policy for Domain Generalization

Tianyi Wu, Liwei Shen, Zhen Dong, Xin Peng, Wenyun Zhao

Primary Area: reinforcement learning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: programmatic policy, reinforcement learning

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Abstract:

Deep reinforcement learning has effectively addressed numerous complex control tasks.
However, when the environment undergoes changes, such as increasing the number of discs from three to four in the `Tower of Hanoi', learned policies often struggle with generalization.
We propose an algorithm for learning programmatic policies capable of capturing environment variations.
In doing so, these policies gain the capability to generalize to instances where certain aspects of the domain exhibit variations, a property we term domain generalization.
We design a Domain Specific Language to construct the structure of the policy.
Through sampling tasks from a task distribution, we can train the policy with a meta-learning algorithm.
Furthermore, our approach incorporates Recurrent Neural Network (RNN) into the structure of the programmatic policy to enhance agent-environment interactions.
Experiment results demonstrate the efficiency of our approach across three environments with domain generalization.
In addition, the learned policy shows its ability to generalize to tasks under different variations of environments.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 7909