Predicates and Fields

The heart of an ORM is defining the mapping between the predicates and Python objects. In Clorm this is acheived by sub-classing the Predicate class and specifying fields that map to the ASP predicate parameters.

The Basics

It is easiest to explain this mapping by way of a simple example. Consider the following ground atoms for the predicates address/2 and pets/2. This specifies that the address of the entity dave is "UNSW Sydney" and dave has 1 pet.

address(dave, "UNSW Sydney").
pets(dave, 1).

Note

A note on ASP syntax. All predicates must start with a lower-case letter and consist of only alphanumeric characters (and underscore). ASP supports three basic types of terms (i.e., the parameters of a predicate); a constant, a string, and an integer. Like the predicate names, constants consist of only alphanumeric characters (and underscore) with a starting lower-case character. This is different to a string, which is quoted and can contain arbitrary characters including spaces.

ASP syntax also supports complex terms (also called functions but we will avoid this usage to prevent confusion with Python functions) which we will discuss later. Note, however that ASP does not support real number values.

To provide a mapping that satisfies the above predicate we need to sub-class the Predicate class and use the ConstantField and StringField classes. These field classes, including the IntegerField, are all sub-classes of the base RawField class.

from clorm import Predicate, ConstantField, StringField, IntegerField

class Address(Predicate):
   entity = ConstantField
   details = StringField

class Pets(Predicate):
   entity = ConstantField
   num = IntegerField(default=0)

Typically, when instantiating a predicate all field values must be provided. The exception is when the field has been defined with a default value, such as with the above definition for the num field of the Pets. So, with the above class definitions we can instantiate some objects:

fact1 = Address(entity="bob", details="Sydney uni")
fact2 = Pets(entity="bob")
fact3 = Pets(entity="bill", num=2)

These object correspond to the following ASP ground atoms (i.e., facts):

address(bob, "Sydney uni").
pets(bob, 0).
pets(bill, 2).

There are some things to note here:

  • Predicate names: ASP uses standard logic-programming syntax, which requires that the names of all predicate/complex-terms must begin with a lower-case letter and can contain only alphanumeric characters or underscore. Unless overriden, Clorm will automatically generate a predicate name for a Predicate sub-class by transforming the class name based on some simple rules:

    • If the first letter is a lower-case character then this is a valid predicate name so the name is left unchanged (e.g., myPredicate => myPredicate).

    • Otherwise, replace any sequence of upper-case only characters that occur at the beginning of the string or immediately after an underscore with lower-case equivalents. The sequence of upper-case characters can include non-alphabetic characters (eg., numbers) and this will still be treated as a single sequence of upper-case characters.

    • The above criteria covers a number of common naming conventions:

      • Snake-case: My_Predicate => my_predicate, MY_Predicate => my_predicate, My_Predicate_1A => my_predicate_1a,

      • Camel-case: MyPredicate => myPredicate, MyPredicate1A => myPredicate1A.

      • Acronym: TCP1 => tcp1.

  • Field order: the order of declared term defintions in the predicate class is important.

  • Field names: besides the Python keywords, Clorm also disallows the following reserved words: raw, meta, clone, Field as these are used as properties or functions of a Predicate object.

  • Constant vs string: In the above example "bob" and "Sydney uni" are both Python strings but because of the entity field is declared as a ConstantField this ensures that the Python string "bob" is treated as an ASP constant. Note, currently it is the users’ responsibility to ensure that the Python string passed to a constant term satisfies the syntactic restriction.

  • The use of a default value: all term types support the specification of a default value.

  • If the specified default is a function then this function will be called (with no arguments) when the predicate/complex-term object is instantiated. This can be used to generated unique ids or a date/time stamp.

Overriding the Predicate Name

As mentioned above, by default the predicate name is calculated from the corresponding class name by transforming the class name to match a number of common naming conventions. However, it is also possible to over-ride the default predicate name with an explicit name.

There are many reasons why you might not want to use the default predicate name mapping. For example, the Python class name that would produce the desired predicate name may already be taken. Alternatively, you might want to distinguish between predicates with the same name but different arities. Note: having predicates with the same name and different arities is a legitimate and common practice with ASP programming.

Overriding the default predicate name requires declaring a Meta sub-class for the predicate definition.

class Address2(Predicate):
   entity = ConstantField
   details = StringField

   class Meta:
       name = "address"

 class Address3(Predicate):
   entity = ConstantField
   details = StringField
   country = StringField

   class Meta:
       name = "address"

Instantiating these classes:

shortaddress = Address2(entity="dave", details="UNSW Sydney")
longaddress = Address3(entity="dave", details="UNSW Sydney", country="AUSTRALIA")

will produce the following matching ASP facts:

address(dave, "UNSW Sydney").
address(dave, "UNSW Sydney", "AUSTRALIA").

Unary Predicates

A unary predicate is a predicate with no parameters and is also a legitimate and reasonable thing to see in an ASP program. Defining a corresponding Python class is straightforward:

class AUnary(Predicate):
    pass

fact = AUnary()

The important thing to note here is that every instantiation of AUnary will correspond to the same ASP fact:

aUnary.

Complex Terms

So far we have shown how to create Python definitions that match predicates with simple terms. However, in ASP it is common to also use complex terms within a predicate, such as:

booking("2018-12-31", location("Sydney", "Australia")).

The Clorm Predicate class definition is able to support the flexiblity required to deal with complex terms. A ComplexTerm class is introduced simply as an alias for the Predicate class.

from clorm import ComplexTerm

class Location(ComplexTerm):
   city = StringField
   country = StringField

The definition for a complex term can be included within a new Predicate definition by using the ComplexTerm.Field property of the ComplexTerm sub-class.

class Booking(Predicate):
    date=StringField
    location=Location.Field

The ComplexTerm.Field property returns a RawField sub-class that is generated automatically when the Predicate sub-class is defined. It provides the functions to automatically convert to, and from, the Predicate sub-class instances and the Clingo symbol objects.

The predicate class containing complex terms can be instantiated in the obvious way:

bk=Booking(date="2018-12-31", location=Location(city="Sydney",country="Australia"))

Note: as with the field definition for simple terms it is possible to specify a complex field definition with default or index parameters. For example, the above Booking class could be replaced with:

class Booking(Predicate):
    date=StringField
    location=Location.Field(index=True,
             default=LocationTuple(city="Sydney", country="Australia"))

Note: From a code readability and conceptual stand point it may be convenient to treat predicates and complex terms as separate classes, however there are cases where this separation breaks down. For example when dealing with the reification of facts there is nothing to be gained by providing two definitions for the predicate and complex term versions of the same non logical term:

p(q(1)).
q(1) :- p(q(1)).

In this example q/1 is both a complex term and predicate and when providing the Python Clorm mapping it is simpler not to separate the two versions:

class Q(Predicate):
   a = IntegerField

class P(Predicate):
   a = Q.Field

Negative Facts

ASP follows standard logic programming syntax and treats the not keyword as default negation (also negation as failure). Using default negation is important to ASP programming as it can lead to more readable and compact modelling of a problem.

However, there may be times when having an explicit notion of negation is also useful, and ASP/Clingo does have support for classical negation; indicated syntactically using the - symbol:

{ a(1..2); b(1..2) }.
-b(N) :- a(N).
-a(N) :- b(N).

The above program chooses amongst the a/1 and b/1 predicates, then for every positive a/1 fact, the corresponding b/1 fact is negated and vice-versa. This will generate nine stable models. For example, if a(2) and b(1) are chosen, then the corresponding negative literals will be -b(2) and -a(1) respectively.

Note: Clingo supports negated literals as well as terms. However, tuples cannot be negated.

f(-g(a)).   % This is valid
f(-(a,b)).  % Error!!!

Clorm supports negation for any fact or term that can be negated by Clingo. Specifying a negative literal simply involves setting sign=False when instantiating the Predicate (or ComplexTerm). Note: unlike the field parameters, the sign parameter must be specified as a named parameter and cannot be specified using positional arguments.

class P(Predicate):
    a = IntegerField

neg_p1 = P(a=1,sign=False)
neg_p1_alt = P(1,sign=False)
assert neg_p1 == neg_p1_alt

Once instantiated, checking whether a fact (or a complex term) is negated can be determined using the sign attribute of Predicate instance.

assert neg_p1.sign == False

Finally, for finer control of the unification process, a Predicate/ComplexTerm can be specified to only unify with either positive or negative facts/terms by setting a sign meta attribute declaration.

class P_pos(Predicate):
    a = IntegerField
    class Meta:
       sign = True
       name = "p"

class P_neg(Predicate):
    a = IntegerField
    class Meta:
       sign = False
       name = "p"

% Instatiating facts
pos_p = P_pos(1)                     % Ok
neg_p_fail = P_pos(1,sign=False)     % throws a ValueError

neg_p = P_neg(1)                     % Ok
pos_p_fail = P_neg(1,sign=False)     % throws a ValueError

% Unifying against raw Clingo positive and negative facts
raws = [Function("p",Number(1)), Function("p",Number(1),positive=False)]
fb = unify([P_pos,P_neg], raw)
assert pos_p in fb
assert neg_p in fb

Field Definitions

Clorm provides a number of standard definitions that specify the mapping between Clingo’s internal representation (some form of Clingo.Symbol) to more natural Python representations. ASP has three simple terms: integer, string, and constant, and Clorm provides three standard definition classes to provide a mapping to these fields: IntegerField, StringField, and ConstantField.

Clorm also provides a SimpleField class that can match to any simple term. This is useful when the parameter of a defined predicate can contain arbitrary simple term types. Clorm takes care of converting the ASP string, constant or integer to a Python string or integer object. Note that both ASP strings and constants are both converted to Python string objects.

In order to convert from a Python string object to an ASP string or constant, SimpleField uses a regular expression to determine if the string matches the pattern of a constant and treats it accordingly. For this reason SimpleField should be used with care in order to ensure expected behaviour, and using the distinct field types is often preferable.

Note

It is worth highlighting that in the above predicate declarations, the field classes do not represent instances of the actual fields. For example, the date string “2018-12-31” is not stored in a StringField object. Rather the field classes provide the implementation of the functions that perform the necessary data conversions. Instantiating a field class in a predicate definition is only necessary to allow options to be specified, such as default values or indexing.

Simple Term Definition Options

There are currently two options when specifying the Python fields for a predicate. We have already seen the default option, but there is also the index option.

Specifying index = True can affect the behaviour when a FactBase container objects are created. While the FactBase class will be discussed in greater detail in the next chapter, here we simply note that it is a convenience container for storing sets of facts. They can be thought of as mini-databases and have some indexing support for improved query performance.

Sub-classing Field Definitions

All field classes inherit from a base class RawField and it’s possible to define arbitrary data conversions by sub-classing RawField. Clorm provides the standard sub-classes StringField, ConstantField, and IntegerField. Clorm also automatically generates an appropriate sub-class for every ComplexTerm definition.

However, it is sometimes also useful to explicitly sub-class the RawField class, or sub-class one of its sub-classes. By sub-classing a sub-class it is possible to form a data conversion chain. To understand why this is useful we consider an example of specifying a date field.

Consider the example of an application that needs a date term for an event tracking application. From the Python code perspective it would be natural to use Python datetime.date objects. However, it then becomes a question of how to encode these Python date objects in ASP (noting that ASP only has three simple term types).

A useful encoding would be to encode a date as a string in YYYYMMDD format (or YYYY-MM-DD for greater readability). Dates encoded in this format satisfy some useful properties such as the comparison operators will produce the expected results (e.g., "20180101" < "20180204"). A string is also preferable to using a similiarly encoded integer value. For example, encoding the date in the same way as an integer would allow incrementing or subtracting a date encoded number, which could lead to unwanted values (e.g., 20180131 + 1 = 20180132 does not correspond to a valid date).

So, adopting a date encoded string we can consider a date based fact for the booking application that simply encodes that there is a New Year’s eve party on the 31st December 2018.

booking("2018-12-31", "NYE party").

Using Clorm this fact can be captured by the following Python Predicate sub-class definition:

from clorm import *

class Booking(Predicate):
   date = StringField
   description = StringField

However, since we encoded the date as simply a StringField it is now up to the user of the Booking class to perform the necessary translations to and from a Python datetime.date objects when necessary. For example:

import datetime
nye = datetime.date(2018, 12, 31)
nyeparty = Booking(date=int(nye.strftime("%Y-%m-%d")), description="NYE Party")

Here the Python nyeparty variable corresponds to the encoded ASP event, with the date term capturing the string encoding of the date.

In the opposite direction to extract the date it is necessary to turn the date encoded string into an actual datetime.date object:

nyedate = datetime.datetime.strptime(str(nyepart.date), "%Y-%m-%d")

The problem with the above code is that the process of creating and using the date in the Booking object is cumbersome and error-prone. You have to remember to make the correct translation both in creating and reading the date. Furthermore the places in the code where these translations are made may be far apart, leading to potential problems when code needs to be refactored.

The solution to this problem is to create a sub-class of RawField that performs the appropriate data conversion. However, sub-classing Rawfield directly requires dealing with raw Clingo Symbol objects. A better alternative is to sub-class the StringField class so you need to only deal with the string to date conversion.

import datetime
from clorm import *

class DateField(StringField):
    pytocl = lambda dt: dt.strftime("%Y-%m-%d")
    cltopy = lambda s: datetime.datetime.strptime(s,"%Y-%m-%d").date()

class Booking(Predicate):
    date=DateField
    description = StringField

The pytocl definition specifies the conversion that takes place in the direction of converting Python data to Clingo data, and cltopy handles the opposite direction. Because the DateField inherits from StringField therefore the pytocl function must output a Python string object. In the opposite direction, cltopy must be passed a Python string object and performs the desired conversion, in this case producing a datetime.date object.

With the newly defined DateField the conversion functions are all captured within the one class definition and interacting with the objects can be done in a more natural manner.

nye = datetime.date(2018,12,31)
nyeparty = Booking(date=nye, description="NYE Party")

print("Event {}: date {} type {}".format(nyeparty, nyeparty.date, type(nyeparty.date)))

will print the expected output:

Event booking(20181231,"NYE Party"): date "2018-12-31" type <class 'datetime.date'>

Note

The pytocl and cltopy functions can potentially be passed bad input. For example when converting a clingo String symbol to a date object the passed string may not correspond to an actual date. In such cases these functions can legitamately throw either a TypeError or a ValueError exception. These exceptions will be treated as a failure to unify when trying to unify clingo symbols to facts. However, any other exception is passed through as a genuine error. This should be kept in mind if you are writing your own field class.

Restricted Sub-class of a Field Definition

Another reason to sub-class a field definition is to restrict the set of values that the field can hold. For example you could have an application where an argument of a predicate is restricted to a specific set of constants, such as the days of the week.

cooking(monday, "Jane"). cooking(tuesday, "Bill"). cooking(wednesday, "Bob").
cooking(thursday, "Anne"). cooking(friday, "Bill").
cooking(saturday, "Jane"). cooking(sunday, "Bob").

When defining a predicate corresponding to cooking/2 it is possible to simply use a ConstantField field for the days.

class Cooking1(Predicate):
   dow = ConstantField
   person = StringField
   class Meta: name = "cooking"

However, this would potentiallly allow for creating erroneous instances that don’t correspond to actual days of the week (for example, with a spelling mistake):

ck = Cooking1(dow="mnday",person="Bob")

In order to avoid these errors it is necessary to subclass the ConstantField in order to restrict the set of values to the desired set. Clorm provides a helper function refine_field() for this use-case. It dynamically defines a new class that restricts the values of an existing field class.

DowField = refine_field("DowField", ConstantField,
   ["sunday","monday","tuesday","wednesday","thursday","friday","saturday"])

class Cooking2(Predicate):
   dow = DowField
   person = StringField
   class Meta: name = "cooking"

try:
   ck = Cooking2(dow="mnday",person="Bob")  # raises a TypeError exception
except TypeError:
   print("Caught exception")

Note

The refine_field() function can also be called with only two arguments, rather than three, by ignoring the name for the generated class. In this case an anonymously generated name will be used.

As well as explictly specifying the set of refinement values, refine_field() also provides a more general approach where a function/functor/lambda can be provided. This function must take a single input and return True if that value is valid for the field. For example, to define a field that accepts only positive integers:

PosIntField = refine_field("PosIntField", NumberField, lambda x : x >= 0)

Finally, it should be highlighted that this mechanism for defining a field restriction works not just for validating the inputs into an ASP program. It can also be used to filter the outputs of the ASP solver as the invalid field values will not unify with the predicate.

For example, in the above program you can separate the cooks on the weekend from the weekday cooks.

WeekendField = refine_field("WeekendField", ConstantField,
   ["sunday","saturday"])
WeekdayField = refine_field("WeekdayField", ConstantField,
   ["monday","tuesday","wednesday","thursday","friday"])

class WeekendCooking(Predicate):
   dow = WeekendField
   person = StringField
   class Meta: name = "cooking"

class WeekdayCooking(Predicate):
   dow = WeekdayField
   person = StringField
   class Meta: name = "cooking"

Using Positional Arguments

So far we have shown how to create Clorm predicate and complex term instances using keyword arguments that match their defined field names, as well as accessing the arguments via the fields as named properties. For example:

from clorm import *

class Contact(Predicate):
    cid=IntegerField
    name=StringField

c1 = Contact(cid=1, name="Bob")

assert c1.cid == 1
assert c1.name == "Bob"

However, Clorm also supports creating and accessing the field data using positional arguments:

c2 = Contact(2,"Bill")

assert c2[0] == 2
assert c2[1] == "Bill"

While Clorm does support the use of positional arguments for predicates, nevertheless it should be used sparingly because it can lead to brittle code that can be hard to debug, and can also be more difficult to refactor as the ASP program changes. However, there are genuine use-cases where it can be convenient to use positional arguments. In particular when defining very simple tuples, where the position of arguments is unlikely to change as the ASP program changes. We discuss Clorm’s support for these cases in the following section.

Working with Tuples

Tuples are a special case of complex terms that often appear in ASP programs. For example:

booking("2018-12-31", ("Sydney", "Australia)).

For Clorm tuples are simply a ComplexTerm sub-class where the name of the corresponding predicate is empty. While this can be set using an is_tuple property of the complex term’s meta class, Clorm also provides specialised support using the more intuitive syntax of a Python tuple. For example, a predicate definition that unifies with the above fact can be defined simply (using the DateField defined earlier):

class Booking(Predicate):
    date=DateField
    location=(StringField,StringField)

Here the location field is defined as a pair of string fields, without having to explictly define a separate ComplexTerm sub-class that corresponds to this pair. To instantiate the Booking class a Python tuple can also be used for the values of location field. For example, the following creates a Boooking instance corresponding to the booking/2 fact above:

bk = Booking(date=datetime.date(2018,12,31), location=("Sydney","Australia"))

While it is unnecessary to define a seperate ComplexTerm sub-class corresponding to the tuple, internally this is in fact exactly what Clorm does. Clorm will transform the above definition into something similar to the following:

class SomeAnonymousName(ComplexTerm):
   city = StringField
   country = StringField
   class Meta:
      is_tuple = True

class Booking(Predicate):
    date=DateField
    location=SomeAnonymousName.Field

Here the ComplexTerm has an internal Meta class with the property is_tuple set to True. This means that the ComplexTerm will be treated as a tuple rather than a complex term with a function name.

One important difference between the implicitly defined and explicitly defined versions of a tuple is that the explicit version allows for field names to be given, while the implicit version will have automatically generated names. However, for simple implicitly defined tuples it would be more common to use positional arguments anyway, so in many cases it can be the preferred alternative. For example:

bk = Booking(date=datetime.date(2018,12,31), location=("Sydney","Australia"))

assert bk.location[0] == "Sydney"

Note

As mentioned previously, using positional arguments is something that should be used sparingly as it can lead to brittle code that is more difficult to refactor. It should mainly be used for cases where the ordering of the fields in the tuple is unlikely to change when the ASP program is refactored.

Debugging Auxiliary Predicates

When integrating an ASP program into a Python based application there will be a set of predicates that are important for inputting a problem instance and outputting a solution. Clorm is intended to provide a clean way of interacting with these predicates.

However, there will typically be other auxiliary predicates that are used as part of the problem formalisation. While they may not be important from the Python application point of view they do become important during the process of developing and debugging the ASP program. During this process it can be cumbersome to build a detailed Clorm predicate definition for each one of these, especially when all you need to do is print the predicate instances to the screen, possibly sorted in some order.

Clorm solves this issue by providing a factory helper function simple_predicate() that returns a Predicate sub-class that will map to any predicate instance with that name and arity.

For example this function could be used for the above booking example if we wanted to extract the booking/2 facts from the model but didn’t care about mapping the data types for the individual parameters. For example to match the ASP fact:

booking("2018-12-31", ("Sydney", "Australia)).

instead of the explicit Booking definition above we could use the simple_predicate() function:

from clorm.clingo import Symbol, Function, String
from clorm import _simple_predicate

Booking_alt = simple_predicate("booking",2)
bk_alt = Booking_alt(String("2018-12-31"), Function("",[String("Sydney"),String("Australia")]))

Note, in this case in order to create these objects within Python it is necessary to use the Clingo functions to explictly create clingo.Symbol objects.

Dealing with Raw Clingo Symbols

As well as supporting simple and complex terms it is sometimes useful to deal with the raw clingo.Symbol objects created through the underlying Clingo Python API.

Raw Clingo Symbols

The Clingo API uses clingo.Symbol objects for dealing with facts; and there are a number of functions for creating the appropriate type of symbol objects (i.e., clingo.Function(), clingo.Number(), clingo.String()).

In essence the Clorm Predicate and ComplexTerm classes simply provide a more convenient and intuitive way of constructing and dealing with these clingo.Symbol objects. In fact the underlying symbols can be accessed using the raw property of a Predicate or ComplexTerm object.

from clorm import *    # Predicate, ConstantField, StringField
from clingo import *   # Function, String

class Address(Predicate):
   entity = ConstantField
   details = StringField

address = Address(entity="dave", details="UNSW Sydney")

raw_address = Function("address", [Function("dave",[]), String("UNSW Sydney")])

assert address.raw == raw_address

Clorm Predicate objects can also be constructed from the raw symbol objects. So assuming the above python code.

address_copy = Address(raw=raw_address)

Note

Not every raw symbol will unify with a given Predicate or ComplexTerm sub-class. If the raw constructor fails to unify a symbol with a predicate definition then a ValueError exception will be raised.

Integrating Clingo Symbols into a Predicate Definition

There are some cases when it might be convenient to combine the simplicity and the structure of the Clorm predicate interface with the flexibility of the underlying Clingo symbol API. For this case it is possible to use the RawField base class itself.

For example when modeling dynamic domains it is often useful to provide a predicate that defines what fluents are true at a given time point, but to allow the fluents themselves to have an arbitrary form.

time(1..5).

true(X,T+1) :- fluent(X), not true(X,T).

fluent(light(on)).
fluent(robotlocation(roby, kitchen)).

true(light(on), 0).
true(robotlocation(roby,kitchen), 0).

In this example instances of the true/2 predicate can have two distinctly different signatures for the first term (i.e., light/1 and robotlocation/2). While the definition of the fluent is important at the ASP level, however, at the Python level we may not be interested in the structure of the fluent, only whether it is true or not. In such a case we can simply treat the fluents themselves as raw Clingo symbol objects.

from clorm import *

class True(Predicate):
   fluent = RawField
   time = IntegerField

Accessing the value of the fluent simply returns the raw Clingo symbol. Also the RawField has the useful property that it will unify with any Clingo.Symbol object and therefore can be used to capture both the light/1 and robotlocation/2 complex terms.

Combining Field Definitions

The above example is useful for cases where you don’t care about accessing the details of individual fluents and therefore it makes sense to simply treat them as a RawField complex term. However, the question naturally arises what to do if you do want more fine-grained access to these fluents.

There are a few possible solutions to this problem, but one obvious answer is to use a field that combines together multiple fields. Such a combined field could be specified manually by explicitly defining a RawField sub-class. However, to simplify this process the combine_fields() factory function has been provided that will return such a combined sub-class.

With reference to the ASP code of the previous example we could add the following Python integration:

from clorm import Predicate, ComplexTerm, IntegerField, ConstantField combine_fields

class Light(ComplexTerm):
   status=ConstantField

class RobotLocation(ComplexTerm):
   robot=ConstantField
   location=ConstantField
   class Meta: name = "robotlocation"

class True(Predicate):
   fluent = combine_fields("FluentField",[Light,RobotLocation])
   time = IntegerField

The combine_fields() function takes two arguments; the first is an optional field name argument and the second is a list of the sub-fields to combine. Note: when trying to unify a value with a combined field the raw symbol values will be unified with the underlying field definitions in the order that they are listed in the call to combine_fields(). This means that care needs to be taken if the raw symbol values could unify with multiple sub-fields; it will only unify with the first successful sub-field. In the above example this is not a problem as the two fluent field definitions do not overlap.

Dealing with Nested Lists

ASP does not have an explicit representation for lists. However a common convention for encoding lists is using a nesting of head-tail pairs; where the head of the pair is the element of the list and the tail is the remainder of the list, being another pair or an empty tuple to indicate the end of the list.

For example encoding a list of “nodes” [1,2,c] for some predicate p, might take the form:

p(nodes,(1,(2,(c,())))).

While, such an encoding can be problematic and can lead to a grounding blowout, nevertheless when used with care can be very useful.

Unfortunately, getting facts containing these sorts of nested lists into and out of Clingo can be very cumbersome. To help support this type of encoding Clorm provides the define_nested_list_field() function. This factory function takes an element field class, as well as an optional new class name, and returns a newly created RawField sub-class that can be used to convert to and from a list of elements of that field class.

from clorm import Predicate, ConstantField, SimpleField, \
   define_nested_list_field

SNLField=define_nested_list("SNLField",SimpleField)

class P(Predicate):
   param=ConstantField
   alist=SNLField

p = P("nodes",[1,2,"c"])
assert str(p) == "p(nodes,(1,(2,(c,()))))"

p = P("nodes",[1,2,"c","A string"])
assert str(p) == '''p(nodes,(1,(2,(c,("A string",())))))'''