python-sat 1.8.dev25__cp312-cp312-manylinux_2_24_x86_64.manylinux_2_28_x86_64.whl

pysat/engines.py

@@ -0,0 +1,1302 @@
+#!/usr/bin/env python
+#-*- coding:utf-8 -*-
+##
+## engines.py
+##
+##  Created on: Sep 20, 2023
+##      Author: Alexey Ignatiev, Zi Li Tan
+##      E-mail: alexey.ignatiev@monash.edu, ztan0050@student.monash.edu
+##
+
+"""
+    ===============
+    List of classes
+    ===============
+
+    .. autosummary::
+        :nosignatures:
+
+        Propagator
+        BooleanEngine
+        LinearConstraint
+        ParityConstraint
+
+    ==================
+    Module description
+    ==================
+
+    This module provides a user with the possibility to define their own
+    propagation engines, i.e. constraint propagators, attachable to a SAT
+    solver. The implementation of this functionality builds on the use of the
+    IPASIR-UP interface [1]_. This may come in handy when it is beneficial to
+    reason over non-clausal constraints, for example, in the settings of
+    satisfiability modulo theories (SMT), constraint programming (CP) and lazy
+    clause generation (LCG).
+
+    .. [1] Katalin Fazekas, Aina Niemetz, Mathias Preiner, Markus Kirchweger,
+        Stefan Szeider, Armin Biere. *IPASIR-UP: User Propagators for CDCL*.
+        SAT. 2023. pp. 8:1-8:13
+
+    .. note::
+
+        Currently, the only SAT solver available in PySAT supporting the
+        interface is CaDiCaL 1.9.5.
+
+    The interface allows a user to attach a single reasoning engine to the
+    solver. This means that if one needs to support multiple kinds of
+    constraints simultaneously, the implementation of the engine may need to
+    be sophisticated enough to make it work.
+
+    It is imperative that any propagator a user defines must inherit the
+    interface of the abstract class :class:`Propagator` and defines all the
+    required methods for the correct operation of the engine.
+
+    An example propagator is shown in the class :class:`BooleanEngine`. It
+    currently supports two kinds of example constraints: linear (cardinality
+    and pseudo-Boolean) constraints and parity (exclusive OR, XOR)
+    constraints. The engine can run in the *adaptive mode*, i.e. it can enable
+    and disable itself on the fly.
+
+    Once an engine is implemented, it should be attached to a solver object by
+    calling :meth:`connect_propagator` of :class:`Cadical195`. The propagator
+    will then need to inform the solver what variable it requires to observe.
+
+    .. code-block:: python
+
+            solver = Solver(name='cadical195', bootstrap_with=some_formula)
+
+            engine = MyPowerfulEngine(...)
+            solver.connect_propagator(engine)
+
+            # attached propagator wants to observe these variables
+            for var in range(some_variables):
+                solver.observe(var)
+
+            ...
+
+    .. note::
+
+        A user is encouraged to examine the source code of
+        :class:`BooleanEngine` in order to see how an external reasoning
+        engine can be implemented and attached to CaDiCaL 1.9.5. Also consult
+        the implementation of the corresponding methods of
+        :class:`.Cadical195`.
+
+    ==============
+    Module details
+    ==============
+"""
+
+#
+#==============================================================================
+from collections import Counter, defaultdict
+import functools
+import itertools
+from typing import List
+
+
+#
+#==============================================================================
+class LinearConstraint:
+    """
+        A possible implementation of linear constraints over Boolean
+        variables, including cardinality and pseudo-Boolean constraints. Each
+        such constraint is meant to be in the less-than form, i.e. a user
+        should transform the literals, weights and the right-hand side of the
+        constraint into this form before creating an object of
+        :class:`LinearConstraint`. The class is designed to work with
+        :class:`BooleanEngine`.
+
+        The implementation of linear constraint propagation builds on the use
+        of counters. Basically, each time a literal is assigned to a positive
+        value, it is assumed to contribute to the total weight on the
+        left-hand side of the constraint, which is calculated and compared to
+        the right-hand side.
+
+        The constructor receives three arguments: ``lits``, ``weights``, and
+        ``bound``. Argument ``lits`` represents a list of literals on the
+        left-hand side of the constraint while argument ``weights`` contains
+        either a list of their weights or a dictionary mapping literals to
+        weights. Finally, argument ``bound`` is the right-hand side of the
+        constraint.
+
+        Note that if no weights are provided, each occurrence of a literal is
+        assumed to have weight 1.
+
+        .. note::
+
+            All weights are supposed to be non-negative values.
+
+        :param lits: list of literals (left-hand side)
+        :param weights: weights of the literals
+        :param bound: right-hand side of the constraint
+
+        :type lits:
+        :type weights: list or dict
+        :type bound: int or float
+    """
+
+    def __init__(self, lits=[], weights={}, bound=1):
+        """
+            Constraint initialiser.
+        """
+
+        if not weights or type(weights) == dict:
+            # the same literal may appear multiple times in the list
+            # so we need to calculate it's actual weight
+            cntr = Counter(lits)
+
+            # left-hand side in terms of literals and their weights
+            self.lits = sorted(cntr.keys())
+            self.wght = {l: cntr[l] * (weights[l] if weights else 1) for l in self.lits}
+        else:
+            assert len(lits) == len(weights), 'The number of weights differs from the number of literals'
+
+            # first, computing the weights (there may be literal repetition)
+            self.wght = defaultdict(lambda: 0)
+            for l, w in zip(lits, weights):
+                self.wght[l] += w
+
+            # second, extracting the list of unique literals
+            self.lits = sorted(self.wght.keys())
+
+        # right-hand side and value counter
+        self.rbnd = bound
+        self.vcnt = 0
+
+        # auxiliary data used during propagation
+        # None is needed here for propagation without an assignment
+        self.lset = set(self.lits + [None])
+        self.lval = None
+
+        # a model that falsifies the constraint (if any)
+        self.fmod = None
+
+        if min(self.wght.values()) == max(self.wght.values()):
+            if self.wght[self.lits[0]] != 1:
+                self.rbnd //= self.wght[self.lits[0]]
+                self.wght = {l: 1 for l in lits}
+
+            # unweighted / cardinality propagation
+            self.propagate = self.propagate_unweighted
+
+            # constraint propagates all literals at once,
+            # i.e. they all share the same reason
+            self.expl = []
+            self.justify = self.justify_unweighted
+            self.abandon = self.abandon_unweighted
+        else:
+            # weighted / pseudo-Boolean propagation
+            self.propagate = self.propagate_weighted
+
+            # constraint may propagate different literals
+            # at different points in time, reasons may differ
+            self.expl = defaultdict(lambda: [])
+            self.justify = self.justify_weighted
+            self.abandon = self.abandon_weighted
+
+        # for weighted case only: a flag for indicating whether we can propagate
+        self.done = False
+
+        # adding literal 'None' with 0-weight
+        self.wght[None] = 0
+
+    def register_watched(self, to_watch):
+        """
+            Add self to the centralised watched literals lists in
+            :class:`BooleanEngine`.
+        """
+
+        for lit in self.lits:
+            to_watch[lit].append(self)
+
+    def attach_values(self, values):
+        """
+            Give the constraint access to centralised values exposed from
+            :class:`BooleanEngine`.
+        """
+
+        self.lval = values
+
+    def propagate_unweighted(self, lit=None):
+        """
+            Get all the consequences of a given literal in the unweighted
+            case. The implementation *counts* how many literals on the
+            left-hand side are assigned to true.
+        """
+
+        if not self.expl:
+            self.vcnt += self.wght[lit]
+
+            if self.vcnt == self.rbnd:  # this is when we should propagate
+                iter1, iter2 = itertools.tee(self.lits)
+                self.expl = list(-l for l in iter1 if self.lval[abs(l)] == l)
+                to_entail = list(-l for l in iter2 if self.lval[abs(l)] is None)
+
+                return to_entail
+
+        return []
+
+    def justify_unweighted(self, dummy_lit):
+        """
+            Provide a reason for a literal propagated by this constraint. In
+            the unweighted case, all the literals propagated by this
+            constraint share the same reason.
+        """
+
+        return self.expl
+
+    def abandon_unweighted(self, dummy_lit):
+        """
+            Clear the reason of a given literal.
+        """
+
+        self.expl.clear()
+
+    def propagate_weighted(self, lit=None):
+        """
+            Get all the consequences of a given literal in the weighted case.
+            The implementation counts the weights of all the literals assigned
+            to true and propagates all the other literals (yet unassigned)
+            such that adding their weights to the total sum would exceed the
+            right-hand side of the constraint.
+        """
+
+        if not self.done:
+            self.vcnt += self.wght[lit]
+
+            iter1, iter2, iter3 = itertools.tee(self.lits, 3)
+            expl = list(-l for l in iter1 if self.lval[abs(l)] == l)
+            to_entail = list(-l for l in iter2 if self.lval[abs(l)] is None and self.vcnt + self.wght[l] > self.rbnd and not self.expl[-l])
+
+            # counting the number of unassigned literals to check
+            # whether there is something left to do in the future
+            nunknown = functools.reduce(lambda x, y: x + 1 if self.lval[abs(y)] is not None else x, iter3, 0)
+
+            # are we done with constraint for now?
+            if len(to_entail) == nunknown:
+                self.done = True
+
+            # setting the reason for each of the propagated literals
+            for l in to_entail:
+                self.expl[l] = expl
+
+            return to_entail
+
+        return []
+
+    def justify_weighted(self, lit):
+        """
+            Provide a reason for a literal propagated by this constraint. In
+            the case of weighted constraints, a literal may have a reason
+            different from the other literals propagated by the same
+            constraint.
+        """
+
+        return self.expl[lit]
+
+    def abandon_weighted(self, lit):
+        """
+            Clear the reason of a given literal.
+        """
+
+        self.expl[lit].clear()
+
+    def unassign(self, lit):
+        """
+            Unassign a given literal, which is done by decrementing the
+            literal's contribution to the total sum of the weights of assigned
+            literals.
+        """
+
+        self.vcnt -= self.wght[lit]
+        self.done = False
+
+    def falsified_by(self, model):
+        """
+            Check if the constraint is violated by a given assignment. Upon
+            receiving such an input assignment, the method counts the sum of
+            the weights of all satisfied literals and checks if it exceeds the
+            right-hand side.
+        """
+
+        value = functools.reduce(lambda x, y: x + self.wght[y] if y in self.lset else x, model, 0)
+
+        if value > self.rbnd:
+            self.fmod = model
+            return True
+
+        return False
+
+    def explain_failure(self):
+        """
+            Provide a reason clause for why the previous model falsified
+            the constraint. This will clause will be added to the solver.
+        """
+
+        return [-l for l in self.fmod if l in self.lset]
+
+
+#
+#==============================================================================
+class ParityConstraint:
+    """
+        A possible implementation of parity constraints. These are constraints
+        of the form :math:`l_1 \\oplus l_2 \\oplus \\ldots \\oplus l_n = b`
+        where each :math:`l_i` is a Boolean literal while :math:`b` is a
+        Boolean constant. The class is designed to exemplify the work with
+        :class:`BooleanEngine`.
+
+        The implementation is pretty naive. It propagates a last unassigned
+        literal if all but one literals got their values. The value the
+        propagated literals is assigned to depends on the other values in the
+        constraint as well as on the right-hand side value.
+
+        The constructor receives two arguments: ``lits`` and ``value``.
+        Argument ``lits`` represents a list of literals on the left-hand side
+        of the constraint while argument ``value`` represents the right-hand
+        side. By default, ``value`` equals ``1``.
+
+        :param lits: list of literals (left-hand side)
+        :param value: right-hand side of the constraint
+
+        :type lits:
+        :type value: bool
+    """
+
+    def __init__(self, lits=[], value=1):
+        """
+            Constraint initialiser.
+        """
+
+        # checking whether the right-hand side is correct
+        assert value in (0, 1), 'incorrect right-hand side value'
+
+        # left-hand side of the parity constraint
+        # trivially removing all the negative literals
+        self.lits = []
+        for lit in lits:
+            if lit < 0:
+                lit = -lit
+                value ^= value
+            self.lits.append(lit)
+
+        # right-hand side, current value, and variable counter
+        self.rval = value
+        self.curr = 0
+        self.vcnt = 0
+
+        # auxiliary data used during propagation
+        # None is needed here for propagation without an assignment
+        self.lset = set(self.lits + [None])
+        self.lval = None
+
+        # reason for the propagated literals
+        self.expl = []
+
+        # a model that falsifies the constraint (if any)
+        self.fmod = None
+
+        # a copy of watched literals, watched length, and
+        # a mapping to the index of an opposite literal
+        self.wlst = None
+        self.wopp = defaultdict(lambda: None)
+
+        # if it is unit clause then propagate its sole literal at level 0
+        self.zlev = self.lits[0] * (2 * self.rval - 1) if len(self.lits) == 1 else None
+
+    def register_watched(self, to_watch):
+        """
+            Add self to the centralised watched literals lists in
+            :class:`BooleanEngine`.
+        """
+
+        if len(self.lits) > 1:
+            # copying the watched lists
+            if self.wlst is None:
+                self.wlst = to_watch
+
+            # we need to trigger this constraint
+            # no matter what value a variable gets
+            for lit in self.lits[:2]:
+                self.wopp[+lit] = len(self.wlst[-lit])
+                self.wopp[-lit] = len(self.wlst[+lit])
+
+                self.wlst[+lit].append(self)
+                self.wlst[-lit].append(self)
+
+    def replace_watched(self, old, new):
+        """
+            Register this constraint as to be watched for a new literal.
+        """
+
+        # clean up the old literal
+        p, n = self.wopp[-old], self.wopp[+old]
+        self.wlst[+old][p] = n
+        self.wopp[+old] = None  # technically, we don't have to remove these
+        self.wopp[-old] = None
+
+        # register the new literal
+        self.wopp[+new] = len(self.wlst[-new])
+        self.wopp[-new] = len(self.wlst[+new])
+        self.wlst[+new].append(self)
+        self.wlst[-new].append(self)
+
+    def attach_values(self, values):
+        """
+            Give the constraint access to centralised values exposed from
+            :class:`BooleanEngine`.
+        """
+
+        self.lval = values
+
+    def propagate(self, lit=None):
+        """
+            Get all the consequences of a given literal. Propagation here
+            should be similar to that of normal clauses (2-watched scheme).
+        """
+
+        if not self.expl:
+            if lit:
+                if self.lits[0] in (lit, -lit):  # this is the first literal
+                                                # so make it the second one
+                    self.lits[0], self.lits[1] = self.lits[1], self.lits[0]
+
+                # find the new watch
+                value = self.lval[abs(self.lits[1])] in self.lset
+                for i in range(2, len(self.lits)):
+                    v = abs(self.lits[i])
+
+                    if self.lval[v] is None:  # v is going to be a new watch
+                        self.replace_watched(lit, self.lits[i])
+                        self.lits[1], self.lits[i] = self.lits[i], self.lits[1]
+                        break
+
+                    value ^= self.lval[v] in self.lset
+                else:  # failed to find a new watch!
+                    self.expl = [-self.lval[abs(self.lits[i])] for i in range(1, len(self.lits))]
+                    to_entail = self.lits[0]
+                    if value == self.rval:
+                        to_entail *= -1
+                    return [to_entail]
+            elif self.zlev is not None:
+                self.expl, to_entail, self.zlev = [], self.zlev, None
+                return [to_entail]
+
+        return []
+
+    def justify(self, dummy_lit):
+        """
+            Provide a reason for the literal propagated from here.
+        """
+
+        return self.expl
+
+    def abandon(self, dummy_lit):
+        """
+            Clear the reason of a given literal.
+        """
+
+        self.expl.clear()
+
+    def unassign(self, lit):
+        """
+            Unassign a given literal. A dummy method, which does nothing.
+        """
+
+        pass
+
+    def falsified_by(self, model):
+        """
+            Check if the constraint is violated by a given assignment, which
+            is done by xor'ing up all the values on the left-hand side (given
+            a model) and comparing with the right-hand side value.
+        """
+
+        value = functools.reduce(lambda x, y: x ^ 1 if y in self.lset else x, model, 0)
+
+        if value != self.rval:
+            self.fmod = model
+            return True
+
+        return False
+
+    def explain_failure(self):
+        """
+            Provide a reason clause for why the previous model falsified
+            the constraint. This will clause will be added to the solver.
+        """
+
+        return [-l for l in self.fmod if l in self.lset or -l in self.lset]
+
+
+#
+#==============================================================================
+class Propagator(object):
+    """
+        An abstract class for creating external user-defined propagators /
+        reasoning engines to be used with solver :class:`Cadical195` through
+        the IPASIR-UP interface. All user-defined propagators should inherit
+        the interface of this abstract class, i.e. all the below methods need
+        to be properly defined. The interface is as follows:
+
+        .. code-block:: python
+
+            class Propagator(object):
+                def on_assignment(self, lit: int, fixed: bool = False) -> None:
+                    pass      # receive a new literal assigned by the solver
+
+                def on_new_level(self) -> None:
+                    pass      # get notified about a new decision level
+
+                def on_backtrack(self, to: int) -> None:
+                    pass      # process backtracking to a given level
+
+                def check_model(self, model: List[int]) -> bool:
+                    pass      # check if a given assignment is indeed a model
+
+                def decide(self) -> int:
+                    return 0  # make a decision and (if any) inform the solver
+
+                def propagate(self) -> List[int]:
+                    return [] # propagate and return inferred literals (if any)
+
+                def provide_reason(self, lit: int) -> List[int]:
+                    pass      # explain why a given literal was propagated
+
+                def add_clause(self) -> List[int]:
+                    return [] # add an(y) external clause to the solver
+    """
+
+    def __init__(self):
+        """
+            Initialiser / constructor. Declare/define all the variables
+            required for the propagator's functionality here.
+        """
+
+        # internal flag to declare the propagator "lazy"
+        # (a lazy propagator only checks complete assignments;
+        # currently, checked only when it gets connected)
+        self.is_lazy = False
+
+    def __delete__(self):
+        """
+            Destructor.
+        """
+
+        pass
+
+    def on_assignment(self, lit: int, fixed: bool = False) -> None:
+        """
+            The method is called to notify the propagator about an assignment
+            made for one of the observed variables. An assignment is set to be
+            "fixed" if it is permanent, i.e. the propagator is not allowed to
+            undo it.
+
+            :param lit: assigned literal
+            :param fixed: a flag to mark the assignment as "fixed"
+
+            :type lit: int
+            :type fixed: bool
+        """
+
+        pass
+
+    def on_new_level(self) -> None:
+        """
+            The method called to notify the propagator about a new decision
+            level created by the solver.
+        """
+
+        pass
+
+    def on_backtrack(self, to: int) -> None:
+        """
+            The method for notifying the propagator about backtracking to a
+            given decision level. Accepts a single argument ``to`` signifying
+            the backtrack level.
+
+            :param to: backtrack level
+            :type to: int
+        """
+
+        pass
+
+    def check_model(self, model: List[int]) -> bool:
+        """
+            The method is used for checking if a given (complete) truth
+            assignment satisfies the constraint managed by the propagator.
+            Receives a single argument storing the truth assignment found by
+            the solver.
+
+            .. note::
+
+                If this method returns ``False``, the propagator must be ready
+                to provide an external clause in the following callback.
+
+            :param model: a list of integers representing the current model
+            :type model: iterable(int)
+
+            :rtype: bool
+        """
+
+        pass
+
+    def decide(self) -> int:
+        """
+            This method allows the propagator to influence the decision
+            process. Namely, it is used when the solver asks the propagator
+            for the next decision literal (if any). If the method returns
+            ``0``, the solver will make its own choice.
+
+            :rtype: int
+        """
+
+        return 0
+
+    def propagate(self) -> List[int]:
+        """
+            The method should invoke propagation under the current assignment.
+            It can return either a list of literals propagated or an empty
+            list ``[]``, informing the solver that no propagation is made
+            under the current assignment.
+
+            :rtype: int
+        """
+
+        return []
+
+    def provide_reason(self, lit: int) -> List[int]:
+        """
+            The method is called by the solver when asking the propagator for
+            the reason / antecedent clause for a literal the propagator
+            previously inferred. This clause will be used in the following
+            conflict analysis.
+
+            .. note::
+
+                The clause must contain the propagated literal.
+
+            :param lit: literal to provide reason for
+            :type lit: int
+
+            :rtype: iterable(int)
+        """
+
+        pass
+
+    def add_clause(self) -> List[int]:
+        """
+            The method is called by the solver to add an external clause if
+            there is any. The clause can be arbitrary but if it is
+            root-satisfied or tautological, the solver will ignore it without
+            learning it.
+
+            Root-falsified literals are eagerly removed from the clause.
+            Falsified clauses trigger conflict analysis, propagating clauses
+            trigger propagation. Unit clauses always (unless root-satisfied,
+            see above) trigger backtracking to level 0.
+
+            An empty clause (or root falsified clause, see above) makes the
+            formula unsatisfiable and stops the search immediately.
+
+            :rtype: iterable(int)
+        """
+
+        pass
+
+
+#
+#==============================================================================
+class BooleanEngine(Propagator):
+    """
+        A simple *example* Boolean constraint propagator inheriting from the
+        class :class:`Propagator`. The idea is to exemplify the use of
+        external reasoning engines. The engine should be general enough to
+        support various constraints over Boolean variables.
+
+        .. note::
+
+            Note that this is not meant to be a model implementation of an
+            external engine. One can devise a more efficient implementation
+            with the same functionality.
+
+        The initialiser of of the class object may be provided with a list of
+        constraints, each being a tuple ('type', constraint), as a value for
+        parameter ``bootstrap_with``.
+
+        Currently, there are two types of constraints supported (to be
+        specified) in the constraints passed in: ``'linear'`` and ``'parity'``
+        (exclusive OR). The former will be handled as objects of class
+        :class:`LinearConstraint` while the latter will be transformed into
+        objects of :class:`ParityConstraint`.
+
+        Here, each type of constraint is meant to have a list of literals
+        stored in variable ``.lits``. This is required to set up watched lists
+        properly.
+
+        The second keyword argument ``adaptive`` (set to ``True`` by default)
+        denotes the fact that the engine should check its own efficiency and
+        disable or enable itself on the fly. This functionality is meant to
+        exemplify how adaptive external engines can be created. A user is
+        referred to the source code of the implementation for the details.
+    """
+
+    def __init__(self, bootstrap_with=[], adaptive=True):
+        """
+            Initialiser.
+        """
+
+        self.vars = []                       # all known variables
+        self.vset = set()                    # a set of all variables
+        self.cons = []                       # all known constraints
+        self.lins = []                       # linear constraints
+        self.xors = []                       # parity constraints
+        self.wlst = defaultdict(lambda: [])  # constraints to watch
+
+        # this propagator may disable itself depending on the circumstances
+        self.is_adaptive = adaptive
+
+        # setting up individual constraint handlers
+        if bootstrap_with:
+            for cs in bootstrap_with:
+                cs = self._add_constraint(cs)
+
+                # updating the set of known variables
+                self.vset.update(set(abs(l) for l in cs.lits))
+
+            # setting all known variables
+            self.vars = sorted(self.vset)
+
+            # run the preprocessing techniques (if any)
+            self.preprocess()
+
+            # attaching new constraints after processing to the watched lists
+            for cs in self.cons:
+                cs.register_watched(self.wlst)
+
+        # value and trail handling
+        self.value = {v: None for v in self.vars}
+        self.fixed = {v: False for v in self.vars}
+        self.trail = []
+        self.trlim = []
+        self.props = defaultdict(lambda: [])
+        self.qhead = None
+
+        # decision level
+        self.decision_level = 0
+        self.level = {v: None for v in self.vars}
+
+        # reasons for propagated literals
+        self.origin = defaultdict(lambda: None)
+
+        # this variable points to the constraint previously violated by a
+        # given model offered by the solver
+        self.falsified = None
+
+        # giving constraints access to the values
+        for cs in self.cons:
+            cs.attach_values(self.value)
+
+        # currently, no solver is known to this propagator
+        # this will be update upon calling self.setup_observe()
+        self.solver = None
+
+        # the following constants are used in adaptive mode:
+        # the number of calls to propagate() between model checks and
+        # the number of propagated literals found between those checks
+        # followed by decaying propagation ratio-related constants
+        self.nprops, self.ncalls = 0, 0
+        self.pratio, self.pdecay, self.pbound = 0, 2, 0.2
+
+        # decaying ratio of falsifying assignments
+        # over all assignments offered by the solver
+        self.mratio, self.mdecay, self.mbound = 0, 2, 2
+
+    def adaptive_constants(self, pdecay, pbound, mdecay, mbound):
+        """
+            Set magic numeric constants used in adaptive mode.
+        """
+
+        self.pdecay = pdecay
+        self.pbound = pbound
+
+        self.mdecay = mdecay
+        self.mbound = mbound
+
+    def enable(self):
+        """
+            Notify the solver that the propagator is willing to become active
+            from now on.
+        """
+
+        assert self.solver, 'Solver is not connected!'
+        self.solver.enable_propagator()
+
+    def disable(self):
+        """
+            Notify the solver that the propagator should become inactive as it
+            does not contribute much to the inference process. From now on, it
+            will only be called to check complete models obtained by the
+            solver (see :meth:`check_model`).
+        """
+
+        assert self.solver, 'Solver is not connected!'
+        self.solver.disable_propagator()
+
+    def is_active(self):
+        """
+            Return engine's status. It is deemed active if the method returns
+            ``True`` and passive otherwise.
+        """
+
+        assert self.solver, 'Solver is not connected!'
+        return self.solver.propagator_active()
+
+    def add_constraint(self, constraint):
+        """
+            Add a new constraint to the engine and integrate it to the
+            internal structures, i.e. watched lists. Also, return the newly
+            added constraint to the callee.
+        """
+
+        cs = self._add_constraint(constraint)
+
+        # adding it to the watched lists
+        cs.register_watched(self.wlst)
+
+        # updating the other structures of the propagator
+        for lit in cs.lits:
+            var = abs(lit)
+            if var not in self.vset:
+                # this variable is currently unknown
+                self.vset.add(var)
+                self.value[var] = None
+                self.level[var] = None
+                self.fixed[var] = False
+
+                # letting the solver know we observe this variable
+                self.solver.observe(var)
+
+        # passing values to the constraint
+        cs.attach_values(self.value)
+
+    def _add_constraint(self, cs):
+        """
+            Create a new constraint, add to the list and return for further
+            processing.
+        """
+
+        if cs[0] == 'linear':
+            cs = LinearConstraint(lits=cs[1][0],
+                                  weights={} if len(cs[1]) == 2 else cs[1][2],
+                                  bound=cs[1][1])
+            self.lins.append(cs)
+        elif cs[0] == 'parity':
+            cs = ParityConstraint(lits=cs[1][0], value=cs[1][1])
+            self.xors.append(cs)
+        else:
+            assert 0, 'Unknown type of constraint'
+
+        # adding the newly created constraint
+        # to the list of known constraints
+        self.cons.append(cs)
+
+        return cs
+
+    def setup_observe(self, solver):
+        """
+            Inform the solver about all the variables the engine is interested
+            in. The solver will mark them as observed by the propagator.
+        """
+
+        # saving the solver for the future
+        self.solver = solver
+
+        for var in self.vars:
+            self.solver.observe(var)
+
+    def preprocess(self):
+        """
+            Run some (naive) preprocessing techniques if available for the
+            types of constraints under considerations. Each type of
+            constraints is handled separately of the rest of constraints.
+        """
+
+        if self.lins:
+            self.process_linear()
+
+        if self.xors:
+            self.process_parity()
+
+    def process_linear(self):
+        """
+            Process linear constraints. Here we apply simple pairwise
+            summation of constraints. As the number of result constraints is
+            quadratic, we stop the process as soon as we get 100 new
+            constraints. Also, if a result of the sum is longer than each of
+            the summands, the result constraint is ignored.
+
+            This is trivial procedure is made to illustrate how constraint
+            processing can be done. It can be made dependent on user-specified
+            parameters, e.g. the number of rounds or a numeric value
+            indicating when a pair of constraints should be added and when
+            they should not be added. For consideration in the future.
+        """
+
+        newc = []
+
+        for cs1, cs2 in itertools.combinations(self.lins, 2):
+            seen, wght, remd = set(), defaultdict(lambda: 0), 0
+
+            # traversing the first constraint and handling clashes
+            for l in cs1.lits:
+                if l in cs2.lset:
+                    # the literal appears in both constraints
+                    wght[l] = cs1.wght[l] + cs2.wght[l]
+                elif -l in cs2.lset:
+                    # the literal appears in opposite phases
+                    maxp = max([(l, cs1.wght[l]), (-l, cs2.wght[-l])], key=lambda pair: pair[1])
+                    minp = min([(l, cs1.wght[l]), (-l, cs2.wght[-l])], key=lambda pair: pair[1])
+
+                    if maxp[1] != minp[1]:  # the weights aren't equal
+                        wght[maxp[0]] = maxp[1] - minp[1]
+
+                    # saving the remainder
+                    remd += minp[1]
+                else:
+                    # no clash
+                    wght[l] = cs1.wght[l]
+
+                seen.add(abs(l))
+
+            # traversing the remainder of 2nd constraints; no clashes
+            for l in cs2.lits:
+                if abs(l) in seen:
+                    continue
+                wght[l] = cs2.wght[l]
+
+            # right-hand side of the result constraint
+            rbnd = cs1.rbnd + cs2.rbnd - remd
+
+            if wght and rbnd >= 0:
+                #  left-hand size of the result constraint
+                lits = sorted(wght.keys())
+
+                if len(lits) > len(cs1.lits) and len(lits) > len(cs2.lits):
+                    # the result constraint is not valuable -> ignore it
+                    continue
+
+                # all good, recording the constraint
+                newc.append(['linear', [lits, rbnd, wght]])
+
+                if len(newc) == 100:
+                    break
+
+            elif rbnd < 0:  # negative bound signifies unsatisfiability
+                            # stopping immediately
+                newc = [('linear', [[cs1.lits[0]], 0]), ('linear', [[-cs1.lits[0]], 0])]
+                break
+
+        if newc:  # there are some new constraints to add
+            for cs in newc:
+                # adding constraint to the engine
+                cs = self._add_constraint(cs)
+
+                # updating the set of known variables
+                self.vset.update(set(abs(l) for l in cs.lits))
+
+            # updating all known variables
+            self.vars = sorted(self.vset)
+
+    def process_parity(self):
+        """
+            Process parity/XOR constraints. Basically, this runs Gaussian
+            elimination and see if anything can be derived from it.
+        """
+
+        status = True
+
+        # removing None from all sets
+        for cs in self.xors:
+            cs.lset.remove(None)
+
+        # forward simplification
+        for i, csi in enumerate(self.xors):
+            v = csi.lits[0]  # pivot variable
+
+            # going over all the other constraints and eliminating the pivot
+            for j in range(i + 1, len(self.xors)):
+                csj = self.xors[j]
+
+                if v not in csj.lset: # do nothing if pivot var is not present
+                    continue
+
+                # summing i'th and j'th constraints up
+                csj.lset ^= csi.lset
+                csj.rval ^= csi.rval
+
+                if not csj.lset and csj.rval:
+                    # inconsistency detected
+                    status = False
+
+                    # adding two inconsistent level-0 literals: 1 and -1
+                    csi.lits, csj.lits = [1], [-1]
+                    csi.zlev, csj.zlev = 1, -1
+
+                    break
+
+                # final result
+                csj.lits = sorted(csj.lset)
+
+        if status:
+            # back substitution if no inconsistency was detected
+            vals = {}
+            for i in range(len(self.xors) - 1, -1, -1):
+                cs = self.xors[i]
+
+                lits = []
+                for l in cs.lits:
+                    if l in vals:
+                        cs.rval ^= vals[l]  # substitution
+                    else:
+                        lits.append(l)
+                cs.lits = lits
+
+                if len(cs.lits) == 1:  # checking if substitution propagates
+                    vals[cs.lits[0]] = cs.rval
+                    cs.zlev = cs.lits[0] * (2 * cs.rval - 1)
+
+        # updating the sets of literals and putting None back
+        for cs in self.xors:
+            cs.lset = set(cs.lits + [None])
+
+    def on_assignment(self, lit, fixed):
+        """
+            Update the propagator's state given a new assignment.
+        """
+
+        self.level[abs(lit)] = self.decision_level
+
+        if self.qhead is None:
+            # this is the first assignment we get after backtracking
+            # marking that next propagation should start from here
+            self.qhead = len(self.trail)
+
+        self.trail.append(lit)
+
+        if fixed:
+            self.fixed[abs(lit)] = True
+
+    def on_new_level(self):
+        """
+            Keep track of decision level updates.
+        """
+
+        self.decision_level += 1
+        self.trlim.append(len(self.trail))
+
+    def on_backtrack(self, to):
+        """
+            Cancel all the decisions up to a certain level.
+        """
+
+        # undoing the assignments
+        while len(self.trlim) > to:
+            while len(self.trail) > self.trlim[-1]:
+                lit = self.trail.pop()
+                var = abs(lit)
+
+                if self.value[var] is not None and not self.fixed[var]:
+                    # informing all the constraints
+                    for cs in self.wlst[lit]:
+                        cs.unassign(lit)
+
+                    self.value[var] = None
+
+                self.level[var] = None
+
+                # cleaning all the consequences
+                for l in self.props[lit]:
+                    self.origin[l].abandon(l)
+                    self.origin[l] = None
+                self.props[lit] = []
+
+            self.trlim.pop()
+
+        # updating decision level
+        self.decision_level = to
+
+        # updating queue head
+        self.qhead = None
+
+    def check_model(self, model):
+        """
+            Check if a given model satisfies all the constraints.
+        """
+
+        st = True  # no falsified constraints by default
+
+        # checking all constraints one by one
+        for cs in self.cons:
+            if cs.falsified_by(model):
+                self.falsified = cs
+                st = False
+                break
+
+        if self.is_adaptive:
+            self.adaptive_update(st)
+
+        return st
+
+    def adaptive_update(self, satisfied):
+        """
+            Update adaptive mode: either enable or disable the engine. This
+            depends on the statistics accumulated in the current run and
+            whether or not the previous assignment found by the solver
+            satisfied the constraints.
+        """
+
+        # decaying increment
+        self.pratio = self.pratio / self.pdecay + (self.nprops / self.ncalls if self.ncalls else 0)
+
+        # taking into account whether the current model falsified something
+        self.mratio = self.mratio / self.mdecay + (0 if satisfied else 1)
+
+        if satisfied and self.is_active():
+            # we are currently active and it does not pay off
+            # in terms of propagation; should we slack off a bit?
+            if self.pratio < self.pbound:
+                self.disable()
+        elif not satisfied and not self.is_active():
+            # we are currently passive and there are many conflicts
+            # let's put more effort into propagation to avoid them?
+            if self.mratio > self.mbound:
+                self.enable()
+
+        # resetting the moving average stats
+        self.nprops = self.ncalls = 0
+
+    def propagate(self):
+        """
+            Run the propagator given the current assignment.
+        """
+
+        result = []
+
+        if self.qhead is not None:
+            while self.qhead < len(self.trail):
+                lit = self.trail[self.qhead]
+
+                # if we have just propagated an opposite value,
+                # we stop and expect conflict analysis to be performed
+                if self.origin[-lit] is not None:
+                    break
+
+                self.value[abs(lit)] = lit  # actually setting the value
+
+                watched_holes = []
+                for csid, cs in enumerate(self.wlst[lit]):
+                    # checking the constraint
+                    propagated = cs.propagate(lit)
+
+                    # extracting the reason
+                    for l in propagated:
+                        if self.origin[l] is None:
+                            # this literal is currently unknown
+                            self.origin[l] = cs
+
+                            # augment the list of entailed literals to return
+                            result.append(l)
+
+                            # recording all the dependencies
+                            self.props[lit].append(l)
+
+                    # checking whether or not this cs is still watched
+                    if self.wlst[lit][csid] is None or type(self.wlst[lit][csid]) == int:
+                        # instead of cs, we have None or an integer, i.e. we
+                        # removed cs from the watched list for this literal
+                        watched_holes.append((csid, self.wlst[lit][csid]))
+
+                # cleaning up watches, should be more-or-less cheap,
+                # doing so for +lit but also for -lit if it is affected
+                if watched_holes:
+                    self.cleanup_watched(lit, watched_holes)
+
+                self.qhead += 1
+
+            if self.is_adaptive:
+                # collecting the stats
+                self.nprops += len(result)
+                self.ncalls += 1
+        else:
+            # propagating without an assignment
+            for cs in self.cons:
+                propagated = cs.propagate()
+
+                # extracting the (empty) reason
+                for l in propagated:
+                    if self.origin[l] is None:
+                        # this literal is currently unknown
+                        self.origin[l] = cs
+
+                        # augment the list of entailed literals to return
+                        result.append(l)
+
+        return result
+
+    def cleanup_watched(self, lit, garbage):
+        """
+            Garbage collect holes in the watched list for +lit
+            (and potentially for -lit).
+        """
+
+        otherside = []
+
+        # going backwards, to make it work
+        for i in range(len(garbage) - 1, -1, -1):
+            pid, nid = garbage[i]
+
+            self.wlst[lit][pid] = self.wlst[lit][-1]
+            self.wlst[lit].pop()
+
+            if nid is not None:
+                # this one has a negative counterpart
+                if pid < len(self.wlst[lit]):
+                    # and it's affected
+                    self.wlst[lit][pid].wopp[-lit] = pid
+
+                otherside.append(nid)
+
+        # taking care of the negative counterparts (if those are affected)
+        for nid in sorted(otherside, reverse=True):
+            self.wlst[-lit][nid] = self.wlst[-lit][-1]
+            self.wlst[-lit].pop()
+
+            if nid < len(self.wlst[-lit]):
+                self.wlst[-lit][nid].wopp[lit] = nid
+
+    def provide_reason(self, lit):
+        """
+            Return the reason clause for a given literal.
+        """
+
+        return [lit] + self.origin[lit].justify(lit)
+
+    def add_clause(self):
+        """
+            Extract a new clause to add to the solver if one exists; return an
+            empty clause ``[]`` otherwise.
+        """
+
+        if self.falsified is None:
+            return []
+
+        to_add = self.falsified.explain_failure()
+
+        # cleaning the previous failure (currently being added)
+        # technically, we should clean self.falsified.fmod too
+        # but it should be safe enough not to do so
+        self.falsified = None
+
+        return to_add