In
computer science, program synthesis is the task to construct a
program that
provably satisfies a given high-level
formal specification. In contrast to
program verification, the program is to be constructed rather than given; however, both fields make use of formal proof techniques, and both comprise approaches of different degrees of automation. In contrast to
automatic programming techniques, specifications in program synthesis are usually non-
algorithmic statements in an appropriate
logical calculus.[1]
The primary application of program synthesis is to relieve the programmer of the burden of writing correct, efficient code that satisfies a specification. However, program synthesis also has applications to
superoptimization and inference of
loop invariants.[2]
Origin
During the Summer Institute of Symbolic Logic at Cornell University in 1957,
Alonzo Church defined the problem to synthesize a circuit from mathematical requirements.[3] Even though the work only refers to circuits and not programs, the work is considered to be one of the earliest descriptions of program synthesis and some researchers refer to program synthesis as "Church's Problem". In the 1960s, a similar idea for an "automatic programmer" was explored by researchers in artificial intelligence.[citation needed]
Since then, various research communities considered the problem of program synthesis. Notable works include the 1969 automata-theoretic approach by
Büchi and
Landweber,[4] and the works by
Manna and
Waldinger (c. 1980). The development of modern
high-level programming languages can also be understood as a form of program synthesis.
21st century developments
This section needs expansion with: a more detailed overview of contemporary approaches. You can help by
adding to it. (December 2022)
The early 21st century has seen a surge of practical interest in the idea of program synthesis in the
formal verification community and related fields. Armando Solar-Lezama showed that it is possible to encode program synthesis problems in
Boolean logic and use algorithms for the
Boolean satisfiability problem to automatically find programs.[5]
Syntax-guided synthesis
In 2013, a unified framework for program synthesis problems called Syntax-guided Synthesis (stylized SyGuS) was proposed by researchers at
UPenn,
UC Berkeley, and
MIT.[6] The input to a SyGuS algorithm consists of a logical specification along with a
context-free grammar of expressions that constrains the syntax of valid solutions.[7] For example, to synthesize a function f that returns the maximum of two integers, the logical specification might look like this:
(f(x,y) = x ∨ f(x,y) = y) ∧ f(x,y) ≥ x ∧ f(x,y) ≥ y
and the grammar might be:
<Exp>::= x | y | 0 | 1 | <Exp> + <Exp> | ite(<Cond>, <Exp>, <Exp>)
<Cond>::=<Exp> <= <Exp>
where "ite" stands for "if-then-else". The expression
ite(x <= y, y, x)
would be a valid solution, because it conforms to the grammar and the specification.
From 2014 through 2019, the yearly Syntax-Guided Synthesis Competition (or SyGuS-Comp) compared the different algorithms for program synthesis in a competitive event.[8] The competition used a standardized input format, SyGuS-IF, based on
SMT-Lib 2. For example, the following SyGuS-IF encodes the problem of synthesizing the maximum of two integers (as presented above):
(set-logic LIA)
(synth-fun f ((x Int) (y Int)) Int
((i Int) (c Int) (b Bool))
((i Int (c x y (+ i i) (ite b i i)))
(c Int (0 1))
(b Bool ((<= i i)))))
(declare-var x Int)
(declare-var y Int)
(constraint (>= (f x y) x))
(constraint (>= (f x y) y))
(constraint (or (= (f x y) x) (= (f x y) y)))
(check-synth)
A compliant solver might return the following output:
((define-fun f ((x Int) (y Int)) Int (ite (<= x y) y x)))
The framework of Manna and Waldinger
Non-clausal resolution rules (unifying substitutions not shown)
The framework is presented in a table layout, the columns containing:
A line number ("Nr") for reference purposes
Formulas that already have been established, including axioms and preconditions, ("Assertions")
Formulas still to be proven, including postconditions, ("Goals"),[note 1]
Terms denoting a valid output value ("Program")[note 2]
A justification for the current line ("Origin")
Initially, background knowledge, pre-conditions, and post-conditions are entered into the table. After that, appropriate proof rules are applied manually. The framework has been designed to enhance human readability of intermediate formulas: contrary to
classical resolution, it does not require
clausal normal form, but allows one to reason with formulas of arbitrary structure and containing any junctors ("
non-clausal resolution"). The proof is complete when has been derived in the Goals column, or, equivalently, in the Assertions column. Programs obtained by this approach are guaranteed to satisfy the specification formula started from; in this sense they are correct by construction.[11] Only a minimalist, yet
Turing-complete,[12]purely functional programming language, consisting of conditional, recursion, and arithmetic and other operators[note 3] is supported.
Case studies performed within this framework synthesized algorithms to compute e.g.
division,
remainder,[13]square root,[14]term unification,[15] answers to
relational database queries[16] and several
sorting algorithms.[17][18]
For example, line 55 is obtained by resolving Assertion formulas from 51 and from 52 which both share some common subformula . The resolvent is formed as the disjunction of , with replaced by , and , with replaced by . This resolvent logically follows from the conjunction of and . More generally, and need to have only two unifiable subformulas and , respectively; their resolvent is then formed from and as before, where is the
most general unifier of and . This rule generalizes
resolution of clauses.[19]
Program terms of parent formulas are combined as shown in line 58 to form the output of the resolvent. In the general case, is applied to the latter also. Since the subformula appears in the output, care must be taken to resolve only on subformulas corresponding to
computable properties.
Logical transformations.
For example, can be transformed to ) in Assertions as well as in Goals, since both are equivalent.
Splitting of conjunctive assertions and of disjunctive goals.
An example is shown in lines 11 to 13 of the toy example below.
This rule allows for synthesis of
recursive functions. For a given pre- and postcondition "Given such that , find such that ", and an appropriate user-given
well-ordering of the domain of , it is always sound to add an Assertion "".[20] Resolving with this assertion can introduce a recursive call to in the Program term.
An example is given in Manna, Waldinger (1980), p.108-111, where an algorithm to compute quotient and remainder of two given integers is synthesized, using the well-order defined by (p.110).
Murray has shown these rules to be
complete for
first-order logic.[21]
In 1986, Manna and Waldinger added generalized E-resolution and
paramodulation rules to handle also equality;[22] later, these rules turned out to be incomplete (but nevertheless
sound).[23]
As a toy example, a functional program to compute the maximum of two numbers and can be derived as follows.[citation needed]
Starting from the requirement description "The maximum is larger than or equal to any given number, and is one of the given numbers", the first-order formula is obtained as its formal translation. This formula is to be proved. By reverse
Skolemization,[note 4] the specification in line 10 is obtained, an upper- and lower-case letter denoting a variable and a
Skolem constant, respectively.
After applying a transformation rule for the
distributive law in line 11, the proof goal is a disjunction, and hence can be split into two cases, viz. lines 12 and 13.
Turning to the first case, resolving line 12 with the axiom in line 1 leads to
instantiation of the program variable in line 14. Intuitively, the last conjunct of line 12 prescribes the value that must take in this case. Formally, the non-clausal resolution rule shown in line 57 above is applied to lines 12 and 1, with
p being the common instance x=x of A=A and x=M, obtained by syntactically
unifying the latter formulas,
F[p being true ∧ x=x, obtained from
instantiated line 1 (appropriately padded to make the context F[⋅] around p visible), and
G[p being x ≤ x ∧ y ≤ x ∧ x = x, obtained from instantiated line 12,
yielding
true ∧ false) ∧ (x ≤ x ∧ y ≤ x ∧ true,
which simplifies to .
In a similar way, line 14 yields line 15 and then line 16 by resolution. Also, the second case, in line 13, is handled similarly, yielding eventually line 18.
In a last step, both cases (i.e. lines 16 and 18) are joined, using the resolution rule from line 58; to make that rule applicable, the preparatory step 15→16 was needed. Intuitively, line 18 could be read as "in case , the output is valid (with respect to the original specification), while line 15 says "in case , the output is valid; the step 15→16 established that both cases 16 and 18 are complementary.[note 5] Since both line 16 and 18 comes with a program term, a
conditional expression results in the program column. Since the goal formula has been derived, the proof is done, and the program column of the "" line contains the program.
^The distinction "Assertions" / "Goals" is for convenience only; following the paradigm of
proof by contradiction, a Goal is equivalent to an assertion .
^When and is the Goal formula and the Program term in a line, respectively, then in all cases where holds, is a valid output of the program to be synthesized. This
invariant is maintained by all proof rules. An Assertion formula usually is not associated with a Program term.
^Only the conditional operator (
?:) is supported from the beginning. However, arbitrary new operators and relations can be added by providing their properties as axioms. In the toy example below, only the properties of and that are actually needed in the proof have been axiomatized, in line 1 to 3.
^While ordinary Skolemization preserves satisfiability, reverse Skolemization, i.e. replacing universally quantified variables by functions, preserves validity.
^Axiom 3 was needed for that; in fact, if wasn't a
total order, no maximum could be computed for uncomparable inputs .
References
^Basin, D.; Deville, Y.; Flener, P.; Hamfelt, A.; Fischer Nilsson, J. (2004). "Synthesis of programs in computational logic". In M. Bruynooghe and K.-K. Lau (ed.). Program Development in Computational Logic. LNCS. Vol. 3049. Springer. pp. 30–65.
CiteSeerX10.1.1.62.4976.
^Alonzo Church (1957). "Applications of recursive arithmetic to the problem of circuit synthesis". Summaries of the Summer Institute of Symbolic Logic. 1: 3–50.
^Zohar Manna, Richard Waldinger (Jan 1980). "A Deductive Approach to Program Synthesis". ACM Transactions on Programming Languages and Systems. 2: 90–121.
doi:
10.1145/357084.357090.
S2CID14770735.
^Zohar Manna and Richard Waldinger (Aug 1987). "The Origin of a Binary-Search Paradigm". Science of Computer Programming. 9 (1): 37–83.
doi:
10.1016/0167-6423(87)90025-6.
^Daniele Nardi and Riccardo Rosati (1992). "Deductive Synthesis of Programs for Query Answering". In Kung-Kiu Lau and Tim Clement (ed.). International Workshop on Logic Program Synthesis and Transformation (LOPSTR). Workshops in Computing. Springer. pp. 15–29.
doi:
10.1007/978-1-4471-3560-9_2.
^Jonathan Traugott (1986). "Deductive Synthesis of Sorting Programs". Proceedings of the International Conference on Automated Deduction.
LNCS. Vol. 230. Springer. pp. 641–660.
Zohar Manna, Richard Waldinger (1975). "Knowledge and Reasoning in Program Synthesis". Artificial Intelligence. 6 (2): 175–208.
doi:
10.1016/0004-3702(75)90008-9.