Memory Augmented Policy Optimization for Program Synthesis

Chen Liang, Mohammad Norouzi (Google Brain). arXiv:1807.02322v5 (January 2019).

Note: Source file named “A focus groups study on data sharing…” — filename-content mismatch documented per D008.

Key Contribution

MAPO improves reinforcement learning for weakly supervised program synthesis and semantic parsing by augmenting the policy gradient with a memory buffer of high-reward programs.

Significance

This is the foundation that Neural Program Planner (Biloki, Liang, Lao, ICLR 2019 Workshop) builds upon. MAPO established the baseline for model-free RL applied to structured prediction, which NPP then improved with model-based planning.

Connection to Vault

Part of the neural program synthesis lineage: Neural Symbolic Machines → MAPO → Neural Program Planner. All involve Ni Lao’s research network.