In this world of ever-increasing automation from self-driving cars to personal robots, how far are we in automating the art of writing software? Surprisingly, the problem of automatically learning programs, also known as “program synthesis”, has been an intriguing research area dating back to 1970s . A lot of work then focused on using deductive reasoning in automated theorem provers to derive programs by systematically transforming the program specifications. These techniques were able to automatically synthesize simple programs involving numerical computations and data structure operations, but didn’t quite catch on since often times writing a complete specification of the program proved to be a more daunting task than writing the original program itself in first place. As a result, the interest in program synthesis research started to die down in late 1980s, but the research area has seen a recent resurgence and is now considered one of the hottest areas in software engineering and programming languages community.
What changed from 1980s?
In my opinion, the three biggest advances that have led to this resurgence are: 1) exceptional advances in constraint solving algorithms (such as SAT and SMT solvers), 2) slightly different formulation of the problem (more on it next), and 3) Moore’s law giving us faster and faster compute power. While the traditional techniques focused on learning arbitrary programs from complete specifications, the recent techniques formulate the problem in two key different ways. First, instead of relying on complete specifications which are hard to write, they allow for incomplete and more natural specification mechanisms such as input-output examples, demonstrations of program behavior, partial programs with holes, natural language, etc. and their combinations. Second, instead of considering Turing-complete languages for the hypothesis space of the possible set of programs, they restrict the hypothesis space using a Domain-specific language. This results in not only faster search algorithms since the hypothesis space is more structured, but also results in learning programs that are more understandable and readable.
Isn’t this an undecidable problem?
Yes, in general it is an undecidable problem. Even answering if a program always terminates is an undecidable problem. Here, synthesis techniques are solving a much harder problem of learning a program that not only terminates but also satisfies the properties defined by the specification. For some fragments of language hierarchy such as restricted regular expressions, there are decidable and efficient synthesis algorithms. Even for more expressive languages, techniques based on bounded reasoning and abstract interpretation have proven to be quite successful. Instead of providing guarantees over complete input space (an undecidable problem), these techniques only consider a finite bounded set of inputs with the assumption that if the learnt program behaves correctly on the bounded set of inputs it is very likely to also work for all possible inputs.
What can we synthesize currently?
Even with all the recent advances, the ideal applications for synthesis techniques have been in learning programs that are small (few tens of lines) but are complex and tricky to write manually. Some of the recent applications include synthesizing programs for data extraction and transformation using examples, parsers from interactive labeling of sub-expressions in a program, program refactoring sequences given an input and output program, network policies from scenarios consisting of packet traces and corresponding actions to the packets, optimal platform-tuned implementations of signal processing transforms from their mathematical descriptions, compiler for a low-power spatial architectures that partitions the program into optimal fragments, efficient synchronization in concurrent programs, low-level bitvector algorithms, type completions, hints and feedback for programming assignments, and many many other cool applications . Let me expand a bit more on how synthesis techniques are being applied concretely in two interesting domains of Data Wrangling and Computer-aided Education.
Data Wrangling & Computer-aided Education
Data wrangling is a term coined to refer to the process of converting data in a raw format to a suitable format that allows for subsequent useful analysis. Since writing such data extraction/transformation scripts manually is cumbersome and sometimes even beyond the programming expertise of the users, some studies estimate that this process of data wrangling takes 80% of the total analysis time. The FlashFill system [3,4], often quoted as one of the top new features in Microsoft Excel 2013, allows users to perform regular-expression based data transformations such as splitting and merging by only providing a few input-output examples without the need to write complex Excel macros or VBA. These techniques have been extended to learn semantic data type transformations (such as date, address, etc.) and table join/lookup programs by examples. The FlashExtract system  allows users to perform data extraction from semi-structured text files using few examples. The key idea behind these systems is to first design a domain-specific language that is expressive enough to encode majority of the data wrangling tasks but at the same time is concise enough for amenable learning. These systems then use Version-space algebra based techniques to efficiently search over the large space of programs in the DSL to learn the programs that are consistent with the user-provided examples.
Computer-aided Education is another domain where synthesis techniques are finding an interesting application. There has been a lot of interest recently in making quality education accessible to students worldwide using education initiatives such as EdX, Coursera, and Udacity. The massive open online courses (MOOC) on these platforms are typically taken by hundreds of thousands of students, which presents a massive challenge to provide quality feedback to these students in the same way teachers do in traditional classrooms. Since there are several different ways to solve a given problem especially in programming courses, we cannot simply use a syntactic approach to provide feedback. The AutoProf system  uses synthesis techniques to automatically generate feedback on introductory programming assignments. Given an incorrect student submission, a reference solution, and an error model capturing the set of common mistakes students typically make for a given problem, AutoProf uses the Sketch synthesis system to compute minimal changes to the student program such that it becomes functionally equivalent to the reference solution. It has been used by TAs to grade homework assignments, and is also being integrated on the MITx platform. A similar auto-grading system CPSGrader  was developed for grading laboratory problems in the area of cyber-physical system using parameter synthesis techniques, and was successfully used for grading thousands of submissions for the online edX class. AutomataTutor  is another such system that uses synthesis techniques for automated grading and feedback generation for finite automata constructions, and is currently being used by thousands of students worldwide.
Isn’t Program Synthesis just Machine Learning?
There has been a long ongoing debate on how machine learning is similar/different from program synthesis, since program synthesis techniques (machines) are essentially learning a program from some specification (training data). However, there are some key differences. First, machine learning techniques typically learn a complex higher-order function over a set of finite feature values, whereas program synthesis techniques learn complex structured programs in a Domain-specific language (sometimes recursive and even Turing-complete). This also results in human-understandability of the learnt programs, which can be manually inspected in contrast to complex higher-dimensional functions learnt by machine learning techniques. Second, the synthesis techniques learn from very few examples (often 1), whereas machine learning techniques requires a large amount of data. Finally, synthesis techniques aren’t very robust to noise in the datasets whereas machine learning techniques are able to handle noise quite effectively. A more comprehensive comparison can be found in this nice survey article .
What are the current challenges and trends in Program Synthesis research?
The scalability of synthesis algorithm to learn larger and larger pieces of code has been a perpetual challenge. A recent trend has been to devise new ways to combine machine learning techniques with program synthesis techniques for both scaling the inference process and also making it more robust to handle noise in specification. Another important challenge has been in developing the right user interaction model, where users can provide specification iteratively in an interactive fashion instead of providing the specification in a one-shot black-box manner.
So, are our programming jobs going to be taken up by robots?
Well, not really, at least not in foreseeable future. With the current synthesis algorithms and computational resources, I believe that soon enough we will be able to leverage synthesis technology to automatically write small complex functions using high-level specifications in our day-to-day programming tasks, but the creative burden of coming up with the right design and algorithms for composing these functions to build larger software artifacts would still lie with the programmer. Though, one day we might reach the stage where even the complex algorithms and designs can be automatically generated by the synthesis algorithms. From one of the famous quotes of President John F. Kennedy (slightly rephrased): “If we have the talent to invent new machines that puts people out of work, we have the talent to put them back to work.” Then, we will reinvent programming.
- Zohar Manna, Richard J. Waldinger: Synthesis: Dreams - Programs. IEEE Trans. Software Eng. 5(4): 294-328 (1979)
- Rastislav Bodík, Barbara Jobstmann. Algorithmic program synthesis: introduction. STTT 15(5-6): 397-411 (2013)
- Sumit Gulwani. Automating string processing in spreadsheets using input-output examples. POPL 2011: 317-330
- Sumit Gulwani, William R. Harris, Rishabh Singh: Spreadsheet data manipulation using examples. Communications of the ACM 55(8): 97-105 (2012)
- Vu Le, Sumit Gulwani: FlashExtract: a framework for data extraction by examples. PLDI 2014: 542-553
- Rishabh Singh, Sumit Gulwani, Armando Solar-Lezama: Automated feedback generation for introductory programming assignments. PLDI 2013: 15-26
- Garvit Juniwal, Alexandre Donzé, Jeff C. Jensen, Sanjit A. Seshia: CPSGrader: Synthesizing temporal logic testers for auto-grading an embedded systems laboratory. EMSOFT 2014: 24:1-24:10
- Rajeev Alur, Loris D'Antoni, Sumit Gulwani, Dileep Kini, Mahesh Viswanathan: Automated Grading of DFA Constructions. IJCAI 2013
- Sumit Gulwani, José Hernández-Orallo, Emanuel Kitzelmann, Stephen H. Muggleton, Ute Schmid, Benjamin G. Zorn: Inductive programming meets the real world. Communications of the ACM 58(11): 90-99 (2015)
If you liked this you may also like:
Julig, R.K., "Applying formal software synthesis", in Software, IEEE , vol.10, no.3, pp.11-22, 1993.
Kant, E., "Synthesis of mathematical-modeling software", in Software, IEEE , vol.10, no.3, pp.30-41, 1993.