Experiments with a PCCoder extension

TL;DR

This paper extends the PCCoder system with new functional symbols, increasing training complexity but maintaining synthesis accuracy, demonstrating the potential for more expressive program synthesis in a neural network framework.

Contribution

The paper introduces new functional symbols into the PCCoder DSL, expanding its expressiveness and analyzing the impact on training and synthesis performance.

Findings

Training set size doubled with new symbols

Neural network parameters increased significantly

Synthesis accuracy remained stable on test sets

Abstract

Recent research in synthesis of programs written in some Domain Specific Language (DSL) by means of neural networks from a limited set of inputs-output correspondences such as DeepCoder and its PCCoder reimplementation/optimization proved the efficiency of this kind of approach to automatic program generation in a DSL language that although limited in scope is universal in the sense that programs can be translated to basically any programming language. We experiment with the extension of the DSL of DeepCoder/PCCoder with symbols IFI and IFL which denote functional expressions of the If ramification (test) instruction for types Int and List. We notice an increase (doubling) of the size of the training set, the number of parameters of the trained neural network and of the time spent looking for the program synthesized from limited sets of inputs-output correspondences. The result is positive in the sense of preserving the accuracy of applying synthesis on randomly generated test sets.