# Learning compositional programs with arguments and sampling

arXiv: https://arxiv.org/abs/2109.00619 · HTML: https://mlsanity.com/paper/2109.00619/

## Key facts

- Cited by 2 papers in the map (data through 14 Oct 2026).
- Builds on 9 of its 18 references that are in the map.
- First posted on arXiv 1 Sep 2021.

## Builds on (most cited first, 9 of 9)

- [Neural Turing Machines](https://mlsanity.com/paper/1410.5401/) (2014)
- [Modular Multitask Reinforcement Learning with Policy Sketches](https://mlsanity.com/paper/1611.01796/) (2016)
- [Monte Carlo Tree Search: A Review of Recent Modifications and Applications](https://mlsanity.com/paper/2103.04931/) (2021)
- [Neural Programmer-Interpreters](https://mlsanity.com/paper/1511.06279/) (2015)
- [Monte-Carlo Tree Search as Regularized Policy Optimization](https://mlsanity.com/paper/2007.12509/) (2020)
- [Turning 30: New Ideas in Inductive Logic Programming](https://mlsanity.com/paper/2002.11002/) (2020)
- [Learning Compositional Neural Programs with Recursive Tree Search and Planning](https://mlsanity.com/paper/1905.12941/) (2019)
- [Learning Simple Algorithms from Examples](https://mlsanity.com/paper/1511.07275/) (2015)
- [Learning Compositional Neural Programs for Continuous Control](https://mlsanity.com/paper/2007.13363/) (2020)

## Built on by (most cited first, 2 of 2)

- [Synthesizing explainable counterfactual policies for algorithmic recourse with program synthesis](https://mlsanity.com/paper/2201.07135/) (2022)
- [Reinforcement Learning and Data-Generation for Syntax-Guided Synthesis](https://mlsanity.com/paper/2307.09564/) (2023)

## Abstract (arXiv)

One of the most challenging goals in designing intelligent systems is empowering them with the ability to synthesize programs from data. Namely, given specific requirements in the form of input/output pairs, the goal is to train a machine learning model to discover a program that satisfies those requirements. A recent class of methods exploits combinatorial search procedures and deep learning to learn compositional programs. However, they usually generate only toy programs using a domain-specific language that does not provide any high-level feature, such as function arguments, which reduces their applicability in real-world settings. We extend upon a state of the art model, AlphaNPI, by learning to generate functions that can accept arguments. This improvement will enable us to move closer to real computer programs. Moreover, we investigate employing an Approximate version of Monte Carlo Tree Search (A-MCTS) to speed up convergence. We showcase the potential of our approach by learning the Quicksort algorithm, showing how the ability to deal with arguments is crucial for learning and generalization.