Prototype Description

Used Tools and Libraries

The two most note-worthy libraries (pip-packages) that the prototype uses are:

1. clingo: From the popular python-clingo library, we primarily used the AST (Abstract Syntax Tree) features, which enable us to efficiently parse a logic program.

2. networkx: Which we primarily used to find dependency graphs, one for handling SCCs in non-tight ASP, and one for speeding up the (un)foundeness check of the reduction.

Additionally, for our experiments and regression tests we used the grounders and solvers of clingo and IDLV.

Abstract Description of the Prototype

Generally speaking, the prototype has three main parts, where each part corresponds roughly to a transformer from the clingo AST library. Such a transformer traverses the AST and enables us to edit the program, and retrieve necessary information, needed for grounding the program with Newground. So the three main parts are:

  1. Domain Inference
    • Consists of two transformers.

    • First transformer (newground.term_transformer module): Gets the domain of all facts, and additionally creates datastructures needed for analyzing the SCCs for non-tight ASP.

    • Second transformer (newground.domain_transformer module): Given a domain, a single call computes the next step of the domain (i.e., passing domain to head for each rule). Is called until a fixed point is reached.

  2. Aggregate Rewriting
  3. Applying the Reduction

Transformer 101

A “Transformer” is a class of Clingo’s submodule “ast” (Abstract Syntax Tree). Therefore, the clingo submodule AST defines the syntax and grammar, which define valid ASTs. Such an AST, is a tree G=(V,E), where V is a set of AST objects (e.g., a rule, or a term), and an edge exists, whenever there is a child-parent relationship. To see a rough sketch of such an AST, see the image below, which shows the AST approximation of the program b(1). a(X,2) :- b(X).. Note that the this is not the full AST, but just a rough sketch how the AST looks like.

_images/AST_approximation.svg

Besides the AST definition, the submodule features utility functions and classes, where the Transformer class is one such class. An instantiated transformer performs a tree traversal of the AST, where each node can be changed. Technically, when one implements a transformer, one inherits from the transformer class. This has the benefit that all nodes which are not explicitly handled, are handled automatically in the standard way. If one wants to handle a node type explicitly, one can do this by the “visit” methods. These methods combine the prefix visit with an AST type (e.g. rule). Therefore, the behavior of e.g. a rule can be changed by the method visit_rule. The parameter is the AST-object (node), and the return object is the (possibly) changed AST-Object (node).

The following code snippets give an example how the transformer class works, for a given program p(X) :- a(X), 1 <= #count{Z:a(Z)}. a(1).. The first code-snippet implements the transformer, which prints information about the aggregate:

from clingo.ast import Transformer

class MyTransformer(Transformer):
    def visit_BodyAggregate(self, node):
        self.visit_children(node)

        print(node.function)
        print(node.left_guard)
        print(node.right_guard)
        print(node.elements)

        return node

And the second one calls the transformer:

from clingo.ast import parse_string

if __name__ == "__main__":
    program_string = "p(X) :- a(X), 1 <= #count{Z:a(Z)}. a(1)."
    my_transformer = MyTransformer()
    parse_string(program_string, lambda stm: my_transformer(stm))

Combining the code snippets into one file, and executing it via python, one gets the output:

0
<= 1
None
[ast.BodyAggregateElement([ast.Variable(Location(begin=Position(filename='<string>', line=1, column=27), end=Position(filename='<string>', line=1, column=28)), 'Z')], [ast.Literal(Location(begin=Position(filename='<string>', line=1, column=29), end=Position(filename='<string>', line=1, column=33)), 0, ast.SymbolicAtom(ast.Function(Location(begin=Position(filename='<string>', line=1, column=29), end=Position(filename='<string>', line=1, column=33)), 'a', [ast.Variable(Location(begin=Position(filename='<string>', line=1, column=31), end=Position(filename='<string>', line=1, column=32)), 'Z')], 0)))])]

Where 0 indicates the #count aggregate, <= 1 the left aggregate relation, None the right aggregate relation, and the list corresponds to the aggregate-element Z:a(Z).

Domain Inference

As shortly described above, the domain inference works by first calling the Term-Transformer, and then repeatedly calling the Domain-Transformer, until a fixed-point is reached. In more detail the Term-Transformer generates a variety of data-structures, which lay the groundwork for other transformers.

The most significant data-structure is the domain dictionary, which contains the domain. Every item of the domain dictionary, is a key-value pair, where the key is a string and the value is either a list, or a dict. In general there are three types of keys:

  1. 0_terms: Where the corresponding value defines the whole domain.

2. A predicate p: Then the value is a dict, which corresponds to the position in the arity of the predicate. E.e., for a predicate p(X,Y), the value of domain["p"] is a dict, where e.g. domain["p"][0] specifies the domain of the first position in the predicate.

  1. A rule-variable combination term_rule_<RULE>_variable_<VARIABLE>: The corresponding value is a single-element dict, which defines defines the domain of a specific variable in a specific rule.

The second-most siginificant data-structure is the dependency_graph (and all related DS), which is a networkx-DiGraph. It is ordered in such a way that there is a directed edge ((p,q)) between a predicate p and q iff there is a rule, where p occurs in the positive body and q in the head of the rule. The related data-structures help associate the graph back to the original rules, which is needed in a later step.

Aggregate Rewriting

The aggregate-transformer checks whether an aggregate occurs in a rule. If so, then the selected aggregate rewriting technique is used, for each aggregate in the rule. The individual strategies can be found in the package newground.aggregate_strategies.

Main Transformer (Applying the Reduction)

The main-transformer is the heart of the prototype. In total it operates on a per-rule basis, i.e., for each visit_Rule it is decided, whether the rule shall be rewritten, partly rewritten, or printed as it is. Additionally, the handling of a ground and non-ground rule can be distinguished, where the non-ground rule is the standard case. Then the reduction consists of the three main parts, as specified in the publications:

  1. SAT check (newground.main_transformer_helpers.generate_satisfiability_part module)

  2. Guessing the head (newground.main_transformer_helpers.guess_head_part module)

  3. (Un)foundedness (newground.main_transformer_helpers.generate_foundedness_part module)

Additionally, one has to note that the global-rewritings for the reduction (e.g. :- not sat.) are largely done in the MainTransformer class, with the exception of the Level-Mappings, which are processed in newground.main_transformer_helpers.level_mappings_part module.