`satisfaction problems(cid:3)
`
`Rina Dechter and Daniel Frost
`Department of Information and Computer Science
`University of California, Irvine
`Irvine, California, USA -
`fdechter,frostg@ics.uci.edu
`
`September ,
`
`Abstract
`
`Over the past twenty (cid:12)ve years many backtracking algorithms have
`been developed for constraint satisfaction problems. This survey describes
`the basic backtrack search within the search space framework and then
`presents a number of improvements developed in the past two decades,
`including look-back methods such as backjumping, constraint recording,
`backmarking, and look-ahead methods such as forward checking and dy-
`namic variable ordering.
`
` Introduction
`
`Constraint networks have proven successful in modeling mundane cognitive tasks
`such as vision, language comprehension, default reasoning, and abduction, as
`well as specialized reasoning tasks including diagnosis, design, and temporal and
`spatial reasoning. The constraint paradigm is a generalization of propositional
`logic, in that variables may be assigned values from a set with any number
`of elements, instead of only true and false. Flexibility in the number of
`values can improve the ease and naturalness with which interesting problems
`are modeled.
`The contribution of this paper is a survey of several approaches to solv-
`ing constraint satisfaction problems, focusing on the backtracking algorithm
`and its variants, which form the basis for many constraint solution procedures.
`We provide on a careful exposition of each algorithm, its theoretical underpin-
`nings, and its relationship to similar algorithms. Worst-case bounds on time
`and space usage are developed for each algorithm. In addition to the survey,
`
`(cid:3)This work was partially supported by NSF grant IRI- , Air Force O(cid:14)ce of Scienti(cid:12)c
`Research grant AFOSR F - - - , Rockwell MICRO grants ACM- and - .
`
`
`
`CONFIGIT 1048
`
`1
`
`
`
`the paper makes several original contributions in formulation and analysis of
`the algorithms. In particular the look-back backjumping schemes are given a
`fresh exposition through comparison of the three primary variants, Gashnig’s,
`graph-based and con(cid:13)ict-directed. We show that each of these backjumping
`algorithms is optimal relative to its information gathering process. The com-
`plexity of several schemes as a function of parameters of the constraint-graph are
`explicated. Those include backjumping complexity as a function of the depth
`of the DFS traversal of the constraint graph, learning algorithms as a function
`of the induced-width, and look-ahead methods such as partial-lookahead as a
`function of the size of the cycle-cutset of the constraint graph.
`The remainder of the paper is organized as follows. Section de(cid:12)nes the
`constraint framework and provides an overview of the basic algorithms for solv-
`ing constraint satisfaction problems. In Section we present the backtracking
`algorithm. Sections and survey and analyze look-back methods such as
`backjumping and learning schemes while Section surveys look-ahead methods.
`Finally, in Section we present a brief historical review of the (cid:12)eld. Previous
`surveys on constraint processing as well as on backtracking algorithms can be
`found in [Dec , Mac , Kum , Tsa , KvB ].
`
` The constraint framework
`
`. De(cid:12)nitions
`
`A constraint network or constraint satisfaction problem (CSP) is a set of n
`variables X = fx ; : : : ; xng, a set of value domains Di for each variable xi, and
`a set of constraints or relations. Each value domain is a (cid:12)nite set of values, one
`of which must be assigned to the corresponding variable. A constraint RS over
`S (cid:18) X is a subset of the cartesian product of the domains of the variables in S.
`If S = fxi ; : : : ; xirg, then RS (cid:18) Di (cid:2) (cid:1) (cid:1) (cid:1) (cid:2) Dir. S is called the scope of RS.
`A nogood is a particular assignment of values to a subset of variables which is
`not permitted. In a binary constraint network all constraints are de(cid:12)ned over
`pairs of variables. A constraint graph associates each variable with a node and
`connects any two nodes whose variables appear in the same scope.
`A variable is called instantiated when it is assigned a value from its domain;
`otherwise it is uninstantiated. By xi = a or by (xi; a) we denote that variable xi
`is instantiated with value a from its domain. A partial instantiation or partial
`assignment of a subset of X is a tuple of ordered pairs ((x ; a ); :::;(x i; ai)),
`frequently abbreviated to (a ; ::::; ai) or ~ai when the order of the variables is
`known.
`Let Y and S be sets of variables, and let ~y be an instantiation of the variables
`in Y . We denote by ~yS the tuple consisting of only the components of ~y that
`correspond to the variables in S. A partial instantiation ~y satis(cid:12)es a constraint
`RS i(cid:11) ~yS RS. [Rina, the next sentence is new.] ~y is consistent if ~y satis(cid:12)es all
`constraints RT ; T (cid:18) Y . A consistent partial instantiation is also called a partial
`solution. A solution is an instantiation of all the variables that is consistent.
`
`
`
`2
`
`
`
`T1
`
`T2
`
`T3
`
`T5
`
`T4
`
`Unary constraint
`T = :
`Binary constraints
`T , T: f( : ,: ), ( : , : ), (: , : )
`(: , : ), ( : , : ), ( : ,: )g
`T , T : f(: , : ), ( : , : ), ( : ,: )g
`T, T: f( : ,: ), ( : , : ), (: , : )
`(: , : ), ( : , : ), ( : ,: )g
`T , T: f( : ,: ), ( : , : ), (: , : )g
`T , T: f(: , : ), ( : , : ), ( : ,: )g
`
`Figure : The constraint graph and constraint relations of the scheduling prob-
`lem in Example .
`
`Example . The constraint framework is useful for expressing and solving
`scheduling problems. Consider the problem of scheduling tasks T , T, T ,
`T, T, each of which takes hour to complete. The tasks may start at : ,
`: , or : . Any number of tasks can be executed simultaneously, subject to
`the restrictions that T must start after T , T must start before T and after
`T, T cannot execute at the same time as T or T, and T cannot start at
`: .
`With (cid:12)ve tasks and three time slots, we can model the scheduling problem
`by creating (cid:12)ve variables, one for each task, and giving each variable the domain
`f : , : , : g. Another equally valid approach is to create three variables,
`one for each starting time, and to give each of these variables a domain which
`is the powerset of fT , T, T , T, Tg. Adopting the (cid:12)rst approach, the
`problem’s constraint graph is shown in Figure . The constraint relations are
`shown on the right of the (cid:12)gure.
`
`. Constraint algorithms
`
`Once a problem of interest has been formulated as a constraint satisfaction
`problem, it can be attacked with a general purpose constraint algorithm. Many
`CSP algorithms are based on the principles of search and deduction; more so-
`phisticated algorithms often combine both principles. In this section we brie(cid:13)y
`survey the (cid:12)eld of CSP algorithms.
`
`.. Search - backtracking
`
`The term search is used to represent a large category of algorithms which solve
`problems by guessing an operation to perform or an action to take, possibly
`with the aid of a heuristic [Nil , Pea]. A good guess results in a new state
`that is nearer to a goal. If the operation does not result in progress towards the
`goal (which may not be apparent until later in the search), then the operation
`can be retracted and another guess made.
`
`
`
`3
`
`
`
`For CSPs, search is exempli(cid:12)ed by the backtracking algorithm. Backtracking
`search uses the operation of assigning a value to a variable, so that the current
`partial solution is extended. When no acceptable value can be found, the pre-
`vious assignment is retracted, which is called a backtrack. In the worst case the
`backtracking algorithm requires exponential time in the number of variables,
`but only linear space. The algorithm was (cid:12)rst described more than a century
`ago, and since then has been reintroduced several times [BR].
`
`.. Deduction - constraint propagation
`
`To solve a problem by deduction is to apply reasoning to transform the problem
`into an equivalent but more explicit form.
`In the CSP framework the most
`frequently used type of deduction is known as constraint propagation or con-
`sistency enforcing [Mon, Mac, Fre]. These procedures transform a given
`constraint network into an equivalent yet more explicit one by deducing new
`constraints, tightening existing constraints, and removing values from variable
`domains. In general, a consistency enforcing algorithm will make any partial so-
`lution of a subnetwork extendable to some surrounding network by guaranteeing
`a certain degree of local consistency, de(cid:12)ned as follows.
`A constraint network is -consistent if the values in the domain of each vari-
`able satisfy the network’s unary constraints (that is, constraints which pertain
`to a single variable). A network is k-consistent, k (cid:21) , i(cid:11) given any consistent
`partial instantiation of any k (cid:0) distinct variables, there exists a consistent
`instantiation of any kth additional variable [Fre]. The terms node-, arc-, and
`path-consistency [Mac] correspond to -, -, and -consistency, respectively.
`Given an ordering of the variables, the network is directional k-consistent i(cid:11) any
`subset of k (cid:0) variables is k-consistent relative to variables that succeed the
`k (cid:0) variables in the ordering [DP]. A problem that is k-consistent for all k
`is called globally consistent.
`A variety of algorithms have been developed for enforcing local consistency
`[MF, MH, Coo , VHDT , DP]. For example, arc-consistency algo-
`rithms delete certain values from the domains of certain variables, to ensure
`that each value in the domain of each variable is consistent with at least one
`value in the domain of each other variable. Path-consistency is achieved by
`introducting new constraints or nogoods which disallow certain pairs of values.
`Constraint propagation can be used as a CSP solution procedure, although
`doing so is usually not practical. If global consistency can be enforced, then one
`or more solutions can easily be found in the transformed problem, without back-
`tracking. However, enforcing k-consistency requires in general exponential time
`and exponential space in k [Coo ], and so in practice only local consistency,
`with k (cid:20) , is used.
`In Example , enforcing -consistency on the network will result in the value
`: being removed from the domain of T, since that value is incompatable
`with a unary constraint. Enforcing -consistency will cause several other domain
`values to be removed. For instance, the constraint between T and T means
`that if T is scheduled for : , there is no possible time for T , since it must
`
`
`
`4
`
`
`
`occur before T . Therefore, an arc-consistency algorithm will, among other
`actions, remove : from the domain of T .
`Algorithms that enforce local consistency can be performed as a preprocess-
`ing step in advance of a search algorithm.
`In most cases, backtracking will
`work more e(cid:14)ciently on representations that are as explicit as possible, that is,
`those having a high level of local consistency. The value of the tradeo(cid:11) between
`the e(cid:11)ort spent on pre-processing and the reduced e(cid:11)ort spent on search has
`to be assessed experimentally, and is dependent on the character of the prob-
`lem instance being solved [DM ]. Varying levels of consistency-enforcing can
`also be interleaved with the search process, and doing so is the primary way
`consistency enforcing techniques are incorporated into constraint programming
`languages [JL ].
`
`.. Other constraint algorithms
`
`In addition to backtracking search and constraint propagation, other approaches
`to solving constraint problems include stochastic local search and structure-
`driven algorithms. Stochastic methods typically move in a hill-climbing manner
`augmented with random steps in the space of complete instantiations [MJPL ].
`In the CSP community interest in stochastic approaches was sparked by the
`success of the GSAT algorithm [SLM ]. Structure-driven algorithms, which
`employ both search and consistency-enforcing components, emerge from an at-
`tempt to characterize the topology of constraint problems that are tractable.
`Tractable classes of constraint networks are generally recognized by realizing
`that for some problems enforcing low-level consistency (in polynomial time)
`guarantees global consistency. The basic graph structure that supports tractabil-
`ity is a tree [MF]. In particular, enforcing -consistency on a tree-structured
`binary CSP network ensures global consistency along some ordering of the vari-
`ables.
`
` Backtracking
`
`A simple algorithm for solving constraint satisfaction problems is backtracking,
`which traverses the search graph in a depth-(cid:12)rst manner. The order of the
`variables can be (cid:12)xed in advance or determined at run time. The backtracking
`algorithm maintains a partial solution that denotes a state in the algorithm’s
`search space. Backtracking has three phases: a forward phase in which the next
`variable in the ordering is selected; a phase in which the current partial solution
`is extended by assigning a consistent value, if one exists, to the next variable;
`and a backward phase in which, when no consistent value exists for the current
`variable, focus returns to the variable prior to the current variable.
`Figure describes a basic backtracking algorithm. As presented in Figure ,
`the backtracking algorithm returns at most a single solution, but it can easily be
`modi(cid:12)ed to return all solutions, or a desired number. The algorithm employs a
`
`
`
`series of mutable value domains D i such that each D i (cid:18) Di. D i holds the subset
`
`
`
`5
`
`
`
`procedure backtracking
`Input: A constraint network with variables fx ; : : : ; xng and domains
`fD ; : : : ; Dng.
`Output: Either a solution, or noti(cid:12)cation that the network is inconsistent.
`i
`(initialize variable counter)
`
`D i Di
`(copy domain)
`while (cid:20) i (cid:20) n
`instantiate xi selectValue
`if xi is null
`(no value was returned)
`i i (cid:0)
`(backtrack)
`else
`i i +
`
`D i Di
`end while
`if i =
`return \inconsistent"
`else
`return instantiated values of fx ; : : : ; xng
`end procedure
`
`(step forward)
`
`procedure selectValue
`while D
`i is not empty
`
`select an arbitrary element a D i, and remove a from D
`if (xi; a) is consistent with ~ai(cid:0)
`return a
`end while
`return null
`end procedure
`
`(no consistent value)
`
`i
`
`Figure : The backtracking algorithm.
`
`
`
`6
`
`
`
`x
`
`red, blue, green
`
`red, blue
`
`x
`
`x
`
`blue, green
`
`red, green, teal
`
`x
`
`x
`
`red, blue
`
`x
`
`red, blue
`
`blue, green x
`
`Figure : A modi(cid:12)ed coloring problem.
`
`of Di that has not yet been examined under the current partial instantiation.
`The D sets are not needed if the values can be mapped to a contiguous set
`of integers that are always considered in ascending order; in this case a single
`integer can be used as a marker to divide values that have been considered from
`those that have not. We use the D sets to describe backtracking for increased
`generality and to be consistent with the portrayal of more complex algorithms
`later in the paper.
`The selectValue subprocedure is separated from the main backtracking
`routine for clarity. It has access to the current value of all variables in the main
`procedure. The complexity of selectValue depends on how the constraints
`are represented internally. When the CSP is binary, it may be practical to store
`the constraints in a table in contiguous computer memory; a one bit or one byte
`(cid:13)ag denotes whether each possible pair of variables with corresponding values
`is compatible. Accessing an entry in the table is an O( ) operation. Since
`the compatibility of the current candidate assignment (xi; a) must be checked
`with at most n earlier variable-value pairs, the complexity of selectValue for
`binary CSPs can be O(n). With non-binary CSPs, checking compatibility is
`more expensive for two reasons. First, the number of possible constraints is
`exponential in the number of variables, and so the actual constraints are most
`likely stored in a list structure. Checking the list will be O(log c), where c is
`the number of constraints. Another possibility that bears consideration is that
`a constraint can be represented not as data but as a procedure. In this case the
`complexity of selectValue is of course dependent on the complexity of the
`procedures it invokes.
`
`Example . Consider the coloring problem in Figure . The domain of each
`node is written inside the node. Note that not all nodes have the same domain.
`Arcs join nodes that must be assigned di(cid:11)erent colors. Assume backtracking
`search for a solution using two possible orderings: d = x ; x; x ; x; x; x; x
`and d = x ; x; x; x; x; x ; x. The search spaces along orderings d and
`
`
`
`7
`
`
`
`d, as well as those portions explicated by backtracking from left to right, are
`depicted in Figure (a) and (b), respectively. (Only legal states are depicted
`in the (cid:12)gure.)
`
` . Improvements to backtracking
`
`Much of the work in constraint satisfaction during the last decade has been
`devoted to improving the performance of backtracking search. Backtracking
`usually su(cid:11)ers from thrashing, namely, rediscovering the same inconsistencies
`and same partial successes during search. E(cid:14)cient cures for such behavior in
`all cases are unlikely, since the problem is NP-hard [GJ ].
`The performance of backtracking can be improved by reducing the size of
`its expanded search space, which is determined both by the size of the under-
`lying search space, and by the algorithm’s control strategy. The size of the
`underlying search space depends on the way the constraints are represented
`(e.g. on the level of local consistency), the order of variable instantiation, and,
`when one solution su(cid:14)ces, the order in which values are assigned to each vari-
`able. Using these factors, researchers have developed procedures of two types:
`those employed before performing the search, thus bounding the size of the un-
`derlying search space; and those used dynamically during the search and that
`decide which parts of the search space will not be visited. Commonly used pre-
`processing techniques are arc- and path-consistency algorithms, and heuristic
`approaches for determining the variable ordering [HE , Fre, DP ].
`The procedures for dynamically improving the pruning power of backtrack-
`ing can be conveniently classi(cid:12)ed as look-ahead schemes and look-back schemes,
`in accordance with backtracking’s two main phases of going forward to assem-
`ble a solution and going back in case of a dead-end. Look-ahead schemes can
`be invoked whenever the algorithm is preparing to assign a value to the next
`variable. The essence of these schemes is to discover from a restricted amount
`of constraint propagation how the current decisions about variable and value
`selection will restrict future search. Once a certain amount of forward constraint
`propagation is complete the algorithm can use the results to:
`
` . Decide which variable to instantiate next, if the order is not predeter-
`mined. Generally, it is advantageous to (cid:12)rst instantiate variables that
`maximally constrain the rest of the search space. Therefore, the most
`highly constrained variable having the least number of values, is usually
`selected.
`
`. Decide which value to assign to the next variable when there is more than
`one candidate. Generally, when searching for a single solution an attempt
`is made to assign a value that maximizes the number of options available
`for future assignments.
`
`Look-back schemes are invoked when the algorithm is preparing the backtrack-
`ing step after encountering a dead-end. These schemes perform two functions:
`
`
`
`8
`
`
`
`x
`
`1
`
`x
`2
`
`x
`
`3
`
`x
`
`4
`
`b
`
`g
`
`r
`
`r
`
`g
`
`b
`
`5
`
`x x
`
`r t
`6
`
`g
`
`b
`
`b
`
`r
`
`b
`
`g
`
`b
`
`g
`
`b
`
`r
`
`b
`
`r
`
`b
`
`b
`
`b
`
`g
`
`g
`
`b
`
`b
`
`g
`
`r
`
`g
`
`t
`
`r
`
`gt
`
`r
`
`t
`
`r
`
`t
`
`rg
`
`t
`
`r
`
`t
`
`r
`
`t
`
`r
`
`r
`
`r
`
`r
`
`(a)
`
`g
`
`r
`
`b
`
`r
`
`g
`
`r
`
`r
`
`b
`
`r
`
`b
`
`g
`
`t
`
`b
`
`b
`
`t
`
`r
`
`b
`
`(b)
`
`r
`
`b
`
`b
`
`b
`
`r
`
`x
`
`7
`
`1
`
`7
`
`x x
`
`x x x
`
`x x
`
`4
`
`5
`
`6
`
`3
`
`2
`
`Figure : Backtracking search for the orderings (a) d = x ; x; x ; x; x; x; x
`and (b) d = x ; x; x; x; x; x ; x on the example in Figure . Intermediate
`states are indicated by ovals, dead-ends by squares, and solutions by triangles.
`The colors blue, red, green, and teal are abbreviated by their (cid:12)rst letters. Thick
`lines denote portions of the search space explored by backtracking when using
`left-to-right ordering and stopping after the (cid:12)rst solution.
`
`
`
`9
`
`
`
` . Deciding how far to backtrack. By analyzing the reasons for the dead-end,
`irrelevant backtrack points can often be avoided so that the algorithm goes
`back directly to the source of failure, instead of just to the immediately
`preceding variable in the ordering. This procedure is often referred to as
`backjumping.
`
`. Recording the reasons for the dead-end in the form of new constraints, so
`that the same con(cid:13)icts will not arise again later in the search. The terms
`used to describe this function are constraint recording and learning.
`
`In sections and we will describe in detail several principle look-back
`schemes, while section will focus on look-ahead methods.
`
` Backjumping
`
`Backjumping schemes are one of the primary tools for reducing backtracking’s
`unfortunate tendency to rediscover the same dead-ends. A dead-end occurs if
`xi has no consistent values left, in which case the backtracking algorithm will
`go back to xi(cid:0) . Suppose a new value for xi(cid:0) exists but there is no constraint
`between xi and xi(cid:0) . A dead-end will be reached at xi for each value of xi(cid:0) until
`all values of xi(cid:0) have been exhausted. For instance, the problem in Figure will
`have a dead-end at x given the assignment (x = red; x = blue; x = blue; x =
`blue; x = green; x = red). Backtracking will then return to x and instantiate
`it as x = teal, but the same dead-end will be encountered at x. We can (cid:12)nesse
`this situation by identifying the culprit variable responsible for the dead-end and
`then jumping back immediately to reinstantiate the culprit variable, instead of
`repeatedly instantiating the chronologically previous variable. Identi(cid:12)cation of
`a culprit variable in backtracking is based on the notion of con(cid:13)ict sets. For
`ease of exposition, in the following discussion we assume a (cid:12)xed ordering of
`the variables d = (x ; : : : ; xn). This restriction can be lifted without a(cid:11)ecting
`correctness, thus allowing dynamic variable orderings in all the algorithms.
`
`. Con(cid:13)ict sets
`
`A dead-end state at level i indicates that a current partial instantiation ~ai =
`(a ; :::; ai) con(cid:13)icts with xi+ . (a ; :::; ai) is called a dead-end state and xi+ is
`called a dead-end variable. Namely, backtracking generated the consistent tuple
`~ai = (a ; :::; ai) and tried to extend it to the next variable, xi+ , but failed: no
`value of xi+ was consistent with all the values in ~ai.
`The subtuple ~ai(cid:0) = (a ; :::; ai(cid:0) ) may also be in con(cid:13)ict with xi+ , and
`therefore going back to xi and changing its value will not always resolve the
`dead-end at variable xi+ . In general, a tuple ~ai that is a dead-end at level i
`may contain many subtuples that are in con(cid:13)ict with xi+ . Any such partial
`instantiation will not be part of any solution. Backtracking’s control strategy
`often retreats to a subtuple ~aj (alternately, to variable xj) without resolving all
`or even any of these con(cid:13)ict sets. As a result, a dead-end at xi+ is guaranteed
`
`
`
`10
`
`
`
`to recur. Therefore, rather than going to the previous variable, the algorithm
`should jump back from the dead-end state at ~ai = (a ; :::; ai) to the most recent
`variable xb such that ~ab(cid:0) = (a ; :::; ab(cid:0) ) contains no con(cid:13)ict sets of the dead-
`end variable xi+ . As it turns out, identifying this culprit variable is fairly
`easy.
`
`De(cid:12)nition (con(cid:13)ict set) Let ~a = (a ; :::; ai) be a consistent instantiation,
`and let x be a variable not yet instantiated. If no value in the domain of x is
`consistent with ~a, we say that ~a is a con(cid:13)ict set of x, or that ~a con(cid:13)icts with
`variable x. If, in addition, ~a does not contain a subtuple that is in con(cid:13)ict with
`x, ~a is called aminim al con(cid:13)ict set of x.
`
`De(cid:12)nition (i-leaf dead-ends) Given an ordering d = x ; :::; xn, then a tu-
`ple ~ai = (a ; :::; ai) that is consistent but is in con(cid:13)ict with xi+ is called an
`i-leaf dead-end state.
`
`De(cid:12)nition (no-good) Any partial instantiation ~a that does not appear in
`any solution is called ano-good. Minimal no-goods have no no-good subtuples.
`
`A con(cid:13)ict set is clearly a no-good, but there are no-goods that are not con(cid:13)ict
`sets of any single variable. Namely, they may con(cid:13)ict with two or more variables.
`Whenever backjumping discovers a dead-end, it tries to jump as far back
`as possible without skipping potential solutions. These two issues of safety in
`jumping and optimality in the magnitude of a jump need to be de(cid:12)ned relative
`to the information status of a given algorithm. What is safe and optimal for one
`style of backjumping may not be safe and optimal for another, especially if they
`are engaged in di(cid:11)erent levels of information gathering. Next, we will discuss
`two styles of backjumping, Gaschnig’s and graph-based, that lead to di(cid:11)erent
`notions of safety and optimality when jumping back.
`
`De(cid:12)nition (safe jump) Let ~ai = (a ; :::; ai) be an i-leaf dead-end state. We
`say that xj , where j (cid:20) i, is safe if the partial instantiation ~aj = (a ; :::; aj) is a
`no-good, namely, it cannot be extended to a solution.
`
`In other words, we know that if xj’s value is changed no solution will be
`missed.
`
`De(cid:12)nition Let ~ai = (a ; :::; ai) be an i-leaf dead-end. The culprit index
`relative to ~ai is de(cid:12)ned by b = minfj (cid:20) ij ~aj conf licts with xi+ g. We de(cid:12)ne
`the culprit variable of ~ai to be xb.
`
`We use the notions of culprit tuple ~ab and culprit variable xb interchangeably.
`By de(cid:12)nition, ~ab is a con(cid:13)ict set that is minimal relative to pre(cid:12)x tuples, namely,
`those associated with a pre(cid:12)x subset of the ordered variables. We claim that xb is
`both safe and optimal: safe in that ~ab cannot be extended to a solution; optimal
`in that jumping back to an earlier node risks missing a solution. Essentially, if
`the algorithm fails to retract as far back asx b, it is guaranteed to wander in
`the search space rooted at ~ab unnecessarily, but if it retracts further back than
`xb, the algorithm may exclude a portion of the search space in which there is a
`solution.
`
`
`
`11
`
`
`
`Proposition If ~ai is an i-leaf dead-end discovered by backtracking, and xb is
`the culprit variable, then ~ab, is an optimal and safe backjump destination.
`
`Proof: By de(cid:12)nition of a culprit, ~ab is a con(cid:13)ict set of xi+ and therefore
`is a no-good. Consequently, jumping to xb and changing the value ab of xb
`to another consistent value of xb (if one exists) will not result in skipping a
`potential solution. To prove optimality, we need to show that jumping farther
`back to an earlier node risks skipping potential solutions. Speci(cid:12)cally, if the
`algorithm jumps to xb(cid:0)j, then by de(cid:12)nition ~ab(cid:0)j is not a con(cid:13)ict set of xi+ ,
`and therefore it may be part of a solution. Note that ~ab(cid:0)j may be a no-good,
`but the backtracking algorithm cannot determine whether it is without testing
`it further. .
`Computing the culprit variable of ~ai is relatively simple since at most i
`subtuples need to be tested for consistency with xi+ . Moreover, it can be
`computed during search by gathering some basic information while assembling
`~ai. Procedurally, the culprit variable of a dead-end (cid:22)ai = (a ; :::; ai) is the most
`recent variable whose assigned value renders inconsistent the last remaining
`value in the domain of xi+ not ruled out by prior variables.
`Next we present three variants of backjumping. Gaschnig’s backjumping
`implements the idea of jumping back to the culprit variable only at leaf dead-
`ends. Graph-based Backjumping extracts information about irrelevant backtrack
`points exclusively from the constraint graph. Although its strategy for jumping
`back at leaf dead-ends is less than optimal, it introduces the notion of jumping
`back at internal dead-ends as well as leaf dead-ends. Con(cid:13)ict-directed backjump-
`ing combines optimal backjumps at both leaf and internal dead-ends.
`
`. Gaschnig’s backjumping
`
`Rather than wait for a dead-end ~ai to occur, Gaschnig’s backjumping [Gas ]
`records some information while generating ~ai, and uses this information to deter-
`mine the dead-end’s culprit variable xb. The algorithm uses a marking technique
`whereby each variable maintains a pointer to the latest predecessor found in-
`compatible with any of the variable’s values. While generating ~a in the forward
`phase, the algorithm maintains a pointer highi for each variable xi. The pointer
`identi(cid:12)es the latest variable tested for consistency with xi and found to be the
`earliest variable in con(cid:13)ict with a value in D
`i. For example, when no compatible
`values exist for xi and if highi = , the pointer indicates that ~a is a con(cid:13)ict
`set of xi. If ~ai does have a consistent value, then highi is assigned the value
`i (cid:0) . The algorithm jumps from a leaf dead-end ~ai that is inconsistent with
`xi+ , back tox highi+ , its culprit since the dead-end variable is xi+ . Gaschnig’s
`backjumping algorithm is presented in Figure .
`
`Proposition Gaschnig’s backjumping implements only safe and optimal back-
`
`jumps in leaf dead-ends.
`
`Proof: Whenever there is a leaf dead-end ~ai(cid:0) , the algorithm has a partial
`instantiation ~ai(cid:0) = (a ; :::; ai(cid:0) ). Let j = highi. The algorithm jumps back to
`
`
`
`12
`
`
`
`procedure gaschnig’s-backjumping
`Input: A constraint network with variables fx ; : : : ; xng and domains
`fD ; : : : ; Dng.
`Output: Either a solution, or a decision that the network is inconsistent.
`i
`(initialize variable counter)
`
`D i Di
`(copy domain)
`highi
`(initialize pointer to culprit)
`while (cid:20) i (cid:20) n
`instantiate xi selectValue-gbj
`if xi is null
`(no value was returned)
`i highi
`(backjump)
`else
`i i +
`
`D i Di
`highi
`end while
`if i =
`return \inconsistent"
`else
`return instantiated values of fx ; : : : ; xng
`end procedure
`
`select an arbitrary element a D i, and remove a from D
`
`i
`
`procedure selectValue-gbj
`while D
`i is not empty
`
`consistent true
`k
`while k < i and consistent
`if k > highi
`highi k
`if ~ak con(cid:13)icts with (xi; a)
`consistent false
`else
`k k +
`end while
`if consistent
`return a
`end while
`return null
`end procedure
`
`(no consistent value)
`
`Figure : Gaschnig’s backjumping algorithm.
`
`
`
`13
`
`
`
`xj, namely, to the tuple ~aj. Clearly, ~aj is in con(cid:13)ict with xi, so we only have to
`show that ~aj is minimal. Since j = highi when the domain of xi is exhausted,
`and since a dead-end did not happen previously, any earlier ~ak for k < j is not
`a con(cid:13)ict set of xi, and therefore xj is the culprit variable. >From Proposition
` , it follows that this algorithm is safe and optimal.
`
`Example . For the problem in Figure , at the dead-end for x (x =
`red; x = blue; x = blue; x = blue; x = green; x = red), high = ,
`because x = red was ruled out by x = red, blue was ruled out by x = blue,
`and no later variable had to be examined. On returning to x , the algorithm
`
`(cid:12)nds no further values to try (D = ;). Since high = , the next variable
`examined will be x. This demonstrate the algorithm’s ability to backjump on
`leaf dead-ends. On subsequent dead-ends (in x ) it goes back to its preceding
`variable only.
`
`In Gaschnig’s backjumping, a jump happens only at leaf dead-ends. If all
`the children of a node in the search tree lead to dead-ends (as happens with
`x = red in Figure a) the node is termed an internal dead-end. Algorithm
`graph-based backjumping implements jumps at internal dead-ends as well as at
`leaf dead-ends.
`
`. Graph-based backjumping
`
`Graph-based backjumping extracts knowledge about possible con(cid:13)ict sets from
`the constraint graph exclusively. Whenever a dead-end occurs and a s