Sequence generation using a constraint satisfaction problem formulation

[0002] 1. Field of the Invention

[0003] The present invention relates in general to the field of the generation of sequences and, in particular, the generation of sequences of items from a database such that the relationship between the attributes of the component items in the sequence is controlled.

[0004] It is to be understood that in the present description the term <<database>> is used to designate any collection of data, for example covering both pre-stored data and data which is dynamically collected.

[0005] There are many situations in which it is necessary or desirable to create a sequence of items from a collection of items for which data is available, and it is important that the created sequence should be <<coherent>>, that is, that there should be a particular relationship between attributes of the component items in the sequence (typically involving a criterion that the attributes of the component items of the sequence should not be too dissimilar, especially for successive items in the sequence). The sequences in question include temporal sequences and spatial sequences. A typical case where this problem arises is in the field of multimedia and, notably, the problem of automatic generation of recitals (such a recital is an example of a temporal sequence). Here the term <<recital>> is used not only to designate a sequence of musical pieces but, more generally, any temporal sequence of multimedia items (film clips, documentaries, text items, etc.). An example of a problem involving a spatial sequence concerns the designing of a thematically-linked layout of paintings for rooms in an art gallery.

[0006] 2. Description of the Prior Art

[0007] It has been proposed in the paper <<Répresentation de séquences définies sur des ensembles non instanciés par arbre PQR partiel>> by Berkaoui et al [Journées Francophones de Programmation Logique par Contraintes, 1998] to use a Constraint Satisfaction approach for solving scheduling problems. However, the approach proposed in this paper does not consider the <<coherence>> of the sequence of items which is produced, notably it does not enable the similarity or dissimilarity of items making up the sequence to be controlled.

[0008] The present invention provides a system and method which produces <<coherent>> sequences of items in a particular order. The items are, generally, stored in a database and described in terms of data pairs each consisting of an attribute and the corresponding value. The problem of creating the desired sequence is treated as a Constraint Satisfaction Problem (CSP). The sequence to be obtained is specified by formulating a collection of constraints holding on items in the database, each constraint describing a particular property of the sequence. The sequence can be specified by any number of constraints.

[0009] According to the present invention, the items in the database exhibit a particular generic format with associated taxonomies for at least some of the attribute values. Also, the constraints are specified out of a predetermined library of generic constraint classes which have been specially formulated. The special constraint classes used in the present invention allow the expression of desired properties of the target sequence, notably properties of similarity between groups of items, properties of dissimilarity, and properties of cardinality. These constraint classes enable the properties of coherent sequences to be expressed in a particularly simple manner.

[0010] It is the combination of the use of a generic format for items in the database and the special constraint classes which enables the preferred embodiments of the present invention to use CSP solution techniques to solve the difficult combinatorial problem of building an ordered collection of elements satisfying a number of constraints.

[0011] Further features and advantages of the present invention will become clear from the following detailed description of a preferred embodiment thereof, given by way of example, and illustrated in the accompanying drawing.

[0012] FIG. 1 illustrates a sequence of musical items selected from a database according to a method embodying the present invention, together with attribute values thereof.

[0013] The method and system according to the present invention will now be described in detail by reference to a preferred embodiment in which the invention is applied to the automatic composition of musical recitals (that is, musical programs, for example for radio, set-top boxes, etc.). It is to be understood that the present invention is not limited to this application alone.

[0014] The following description of the preferred embodiment of the invention will begin with an explanation of the preferred database format, an indication of how the problem is modelled as a CSP, and a discussion of the special constraint classes used in modelling the recital-composition problem as a CSP, before explaining in detail how the problem can be formulated and solved using constraints in the special constraint classes.

[0015] Each item is described by a collection of attribute/value pairs. In this example, concerning musical recitals, an item describes a musical work and may include the following attributes:

[0016] index (unique identifier of the piece)

[0017] title

[0018] author

[0019] duration

[0020] musical style

[0021] type of leading instrument

[0022] type of orchestration

[0023] etc.

[0024] The values of these attributes may be either discrete, i.e. taken in a fixed predetermined set of possible values (for instance, the type of voice or musical style), or continuous (for instance, the duration). The possible values for certain parameters may require special knowledge (e.g. the taxonomy of musical styles) and so are not discussed here.

[0025] <<Aggregate>> or <<composite>> parameters can also be taken into account, that is, parameters which are a composition of other, more basic parameters. This is a simple extension of the initial principle, and allows constraints to be set on any combination of parameters. For instance, in this example, the database may include the two parameters <<Family>> and <<SubFamily>>, which together make up the notion of <<style>> (e.g. one particular style may be the combination <<Pop>>+<<Rock>>).

[0026] Database items are denoted p_i, where i is the index of the item in the database. The attributes are denoted p_ia_j, where j is the index of the attribute. J is the set of attribute indices. Usually the first attribute is the index of the item itself.

[0027] According to the present invention, collections (notably temporal or spatial sequences) of items can be built by modelling the problem as a CSP and resolving it via constraint satisfaction programming techniques. This is done by representing the collection as a series of constrained variables. Each variable has a domain consisting of the whole set of items in the database. For instance, a sequence of 10 musical pieces is represented by 10 constrained variables v₁, v₂, . . . v₁₀, meaning that the first item is represented by the value of v₁, the second by the value of v₂, etc.

[0028] The coherence of the sequence to be built is expressed by constraints holding on the variables defined above. The creation of the temporal collections is preferably undertaken by a generic solver, which utilizes constraint satisfaction algorithms.

[0029] The constraints necessary to state the properties of the sequence are:

[0030] 1) the standard, simple constraints generally capable of being handled by CSP solution techniques (and already available in the generic solver), such as arithmetic constraints—for example, a constraint imposing that the total duration of the sequence should be equal to (resp. less than, resp. greater than) 1 hour is an instantiation of standard linear arithmetic constraints of the type LinearCombination (resp. LinearInequality), and

[0031] 2) specialized constraints using the specialized constraint classes described below.

[0032] According to the preferred embodiment of the present invention, three specialized constraint classes may be defined, each class being described by a general formula:

[0033] cardinality constraints

[0034] similarity constraints

[0035] dissimilarity constraints

[0036] (the similarity constraints and dissimilarity constraints may be considered to be examples of <<running constraints>>, as explained below).

[0037] CARDINALITY CONSTRAINTS

[0038] Constraints of this class allow properties to be imposed on sets of items. There are two such cardinality constraints, cardinality on items and cardinality on attribute values:

[0039] Cardinality on Items:

[0040] This class of constraints allows it to be stated that the number of items whose attribute j belongs to a given set E must be within [a, b].

[0041] The general formulation is:

[0042] CI (I, j, a, b, E) means that: card {iεI|p_ia_jεE}ε[a,b],

[0043] where I is a set of item indices, j is an attribute index, a and b are integers and E is a subset of the possible values of attribute j.

[0044] For instance, in the context of a recital-composition application, this constraint can be used to state that the number of pieces within a given range (i.e. the first 10 pieces), whose style is <<Rock>>, should be comprised between, say, 4 and 6: CI({1,2, . . . , 10}, s, 4, 6, {<<Rock>>}), where s denotes the index of the parameter corresponding to the style of music of a piece in the database.

[0045] Cardinality on Attribute Values:

[0046] This constraint class allows it to be stated that the number of different values for attribute j of a number of items is within [a,b].

[0047] The general formulation is:

[0048] CA(I j, a, b) means that: card{p_ia_j/iεI}ε[a, b],

[0049] where I is a set of item indices, j is an attribute index, a and b are integers.

[0050] For instance, in the context of a recital-composition application, this constraint can be used to state that the first three pieces should have at least two different tempo (i.e. that they should not all have the same tempo)—CA({1,2,3}, t, 3, 3), where t denotes the index of the parameter corresponding to the tempo in the database.

[0051] Note that cardinality on attribute values can be used to state global difference constraints. The global difference constraint D(I, j), with I a set of indices, and j the index of a parameter, means that: if p_idenotes the value of variable X_ifor every i in I, the following relations hold

[0052] ∀kεI, ∀lεI, p_k.a_j≠p_l.a_j

[0053] In other words, all the variables involved have pairwise different values for attribute j. This constraint can be stated as a particular cardinality constraint:

[0054] CA(I, j, card(I), card(I))

[0055] RUNNING CONSTRAINTS

[0056] The following constraints are termed <<running>> constraints because they are systematically set between pairs of consecutive pieces. There are two classes of running constraint: similarity constraints and dissimilarity constraints.

[0057] SIMILARITY CONSTRAINTS

[0058] This class of constraints allows it to be stated that within a given range successive items are similar to each other. The similarity is defined by a binary predicate holding on one given attribute j. The general formulation is:

[0059] S(a,b,jsimilar(.,.,.)) means that: for each item p_i,i ε[a, b−1], the predicate similar(p_i, p_i+_l, j) is true,

[0060] where a and b are integers representing indices, j is an attribute index, and similar (.,.,.) is a predicate. The first two parameters of the predicate are pieces, and the third one is the index of a parameter.

[0061] For instance, in the context of a recital-composition application, this constraint class allows it to be stated that all pieces in a given contiguous range (say the first 10) should have <<similar styles>>, where the similarity of style is defined by a predicate that associates similar styles by looking up in a predetermined classification (see the <<RecitalComposer>> application discussed below): in this case the <<similar>> predicate is defined by:

[0062] let jbe the index of the style parameter in the database in this case, p.a_jrepresents the style of piece psimilar(piece1,piece2, j):=(piece1.a_jand piece 2.a_jare linked in the taxonomy of styles).

[0063] DISSIMILARITY CONSTRAINTS

[0064] Note that the same constraint class as that discussed above allows it to be stated that all pieces in a given contiguous range (say the first 10) should have different attributes. It suffices to define the <<similar>> predicate by:

[0065] similar (x, y,j):=(x.a_j≠y.a_j).

[0066] The implementations of these special constraint classes do not enforce arc consistency, but ensure that the resolution is sound and complete, that is, all solutions are found, and all solutions are actual solutions, i.e. satisfy all constraints.

[0067] Note that, in all the procedures described below, if the domain of a variable is wiped out by removing values, this means that the corresponding constraint cannot be satisfied in the current state of resolution. Therefore, a failure is raised that is handled by the solver, which triggers a backtrack in order to explore a different part of the search space. This is a classical feature of constraint satisfaction algorithms.

[0068] Implementing a <<Cardinality on Item>> Constraint

[0069] The constraint holds on variables X_i, with i ε I. It requires that there should be between a and b variables that take their values in the set E, where a and b are integers. The implementation of this constraint relies on the following procedure:

[0070] Implementing a <<Cardinality on Value>> Constraint

[0071] The constraint holds on variables X_i, with i ε I. It requires an interval [a,b]. It ensures that the set of values of the variables for a given parameter j, that is, {value(X_l).a_j, . . . , value(X_n).a_j} contains between a and b different elements. The constraint is implemented as follows:

[0072] Implementing a <<Similarity>> Running Constraint

[0073] This constraint S([a,b], j, similar(.,.,.)) is implemented by setting a series of binary constraints on successive variables. The procedure for setting these constraints is the following:

[0074] Where the constraint similar is defined as follows:

[0075] The expression <<similar(x, y, j)>> has a boolean value defined as follows: similar (x, y, j)=true if x.a_jis similar to y.a_j, and the instruction <<breakloop i>> means that the loop labelled i is interrupted.

[0076] Adding redundant constraints to speed up the resolution of cardinality constraints

[0077] When several cardinality constraints are stated over the same set of constrained variables, it can be impossible for the solver to find a solution, or to prove that there is none, within a reasonable period of time. This is, for instance, illustrated by the following example.

[0078] Consider a set {X₁, . . . , X_n} of constrained variables. Consider the following two cardinality constraints:

[0079] 1) CI({I, . . . , 10}, 1, 4, 10, {<<Rock>>, <<Country>>}) where 1 denotes the index of the family parameter, and

[0080] 2) CI({I, . . . , 10}, 1, 4, 10, {<<Pop>>, <<Jazz>>}) where 1 denotes the index of the family parameter.

[0081] This simple problem is hard to solve since the solver will consider each constraint individually, according to the standard model of constraint satisfaction programming. More precisely, standard solvers will instantiate the six first variables with arbitrary chosen values, regardless of the constraints. This is explained by the fact that constraint 1) will have no effects as long as there are more than four variables whose domain contains pieces whose parameter <<family>> (parameter 1 in our case) is either <<Rock>> or <<Country>>. The same argument holds for constraint 2).

[0082] More precisely, the first six variables are instantiated with pieces whose family is <<Funk>>. Constraint 1) will then require that the four remaining variables are instantiated with a piece whose family is either <<Rock>> or <<Country>>. At the same time, the second constraint will require that the four remaining variables are instantiated with a piece whose family is either <<Pop>> or <<Jazz>>. Obviously, these two requirements cannot be met simultaneously.

[0083] Of course, since {<<Rock>>, <<Country>>} and {<<Pop)>>, <<Jazz>>} have no common elements, the solver should handle the constraints as soon as there remain only eight variables whose domain contains pieces whose parameter <<family>>is either <,Rock>>, <<Country>>, <<Pop>> or <<Jazz>>. To overcome this issue, the system according to the preferred embodiment of the invention states additional, redundant constraints, that is, constraints which leave a solution set unchanged but which speed up the resolution of the problem. In the example above, the relevant redundant constraint to add is the following:

[0084] 3) CI({1, . . . , 10}, 1, 8, 10, {<<Rock>>, <<Country>>, <<Pop>>, <<Jazz>>})

[0085] More generally, once a cardinality constraint on items is stated, the following procedure is applied to create additional redundant cardinality constraints:

[0086] In an application of the present invention to the generation of musical recitals (that is, sequences of musical pieces), the items in a sample database of Sony Music musical pieces were described using the following attributes:

[0087] index: an integer

[0088] title: a string of characters

[0089] author: a string of characters

[0090] Family: a family within a predetermined list of families (see below)

[0091] subFamily: a family within a predetermined list of sub-families (see below)

[0092] duration: an integer

[0093] rhythmType: a type of rhythm from within the taxonomy of possible rhythm types

[0094] tempo: a type of tempo from within the taxonomy of possible tempos

[0095] instrumentType1: a type of instrument from within the taxonomy of possible instrument types

[0096] instrumentType2: a type of instrument from within the taxonomy of possible instrument types

[0097] voiceType: a type of voice from within the taxonomy of possible voice types

[0098] The Family and sub-Family in this application were described using a predetermined taxonomy of musical styles which linked them together according to stylistic similarities. This was done by introducing an aggregated attribute, called <<style>>, whose value is the combination of <<Family>> and <<subFamily>>. For instance, the style <<Jazz-Swing>>, is the combination of the Family <<Jazz>> with the subFamily <<Swing>>.

[0099] This taxonomy links styles together. For instance, the style <<Jazz-Swing>> is related in this taxonomy to the style <<Jazz-Crooner>>, the style <<Jazz-Crooner>> is related to the style <<Country-Crooner>>, and so on. This taxonomy naturally yields a predicate similar(x,y), where x and y denote styles.

[0100] In the preferred implementation of the present invention for this application, the automatic recital composer was embodied using a general purpose computer, suitably programmed, and the user was presented with a visual display on a monitor allowing him to specify constraint sets graphically. Solutions could be computed for an arbitrary database and the solution was displayed in a specific window on the monitor display.

[0101] The problem in this example was to automatically generate a sequence of musical pieces, selected in the database, to form a recital which could be entitled <<A gentle path through soul music>>, and meeting the following criteria:

[0102] the sequence should contain 12 pieces (to fit on a CD or Minidisc)

[0103] the sequence should be continuous stylistically: each piece belonging to a style <<similar>> to the style of the preceding piece—this similarity being given by the underlying classification of styles

[0104] start by a <Soul-Jazz>> piece and end by a <<Soul-Crooner>> piece but not going directly from <<Soul-Jazz>> to <<Soul-Crooner>> (so that the system was forced to find a more interesting path)

[0105] preferably start rather slowly (avoiding very fast tempos at the beginning) and end rather quickly (avoiding very slow tempos at the end)

[0106] evolving continuously tempo-wise: each piece having a tempo which is similar to the tempo of the preceding piece—this similarity is given by the underlying hierarchical relation in the classification of tempos

[0107] all pieces should be different (this is usual for a program of this kind of duration)

[0108] all authors should be different.

[0109] The above requirements for the musical sequence may be faithfully represented by formulating a constraint satisfaction problem having the following constraints:

[0110] simple unary constraint on the first and last pieces to set the imposed styles

[0111] a running constraint of similarity of styles

[0112] simple running incompatibleValue constraint on styles

[0113] simple unary constraint on tempo for the first and last pieces running similarity constraint on tempos

[0114] global difference constraint on pieces (index attribute)

[0115] global difference constraint on authors.

[0116] More technically, these constraints are stated as follows

[0117] CI({1}, style, 1, 1, {<<Soul-Jazz>>})

[0118] CI({1}, style, 1, 1, {<<Soul-Crooner>>})

[0119] S([1, 12], style, similar (.,.,.))

[0120] S([1, 12], style, different (.,.,.))

[0121] CI({1}, tempo, 0, 0, {<<fast, very fast>>})

[0122] CI({1}, tempo, 0, 0, {<<slow, very slow>>})

[0123] S([1, 12], tempo, similar (.,.,.))

[0124] allDiff ([1,12], index)

[0125] allDiff ([1, 12], authors) Note that for ease of comprehension the parameters have been denoted by their names rather than by their index.

[0126] The result of application of the present invention to the above problem was to generate the sequence of musical pieces represented in FIG. 1. In the table of FIG. 1, the titles, authors, styles, durations, tempos, voice types and Instrument types (1 and 2) of the generated sequence of musical pieces are indicated.

[0127] It will be seen that application of the method of the present invention to a particular problem of automatic recital composition gave rise to a solution which met the specified design criteria.

[0128] The technique described above for automatic generation of recitals can, for example, be implemented in a module for incorporation into music-playing devices such as those having means for interacting arbitrarily with any of a plurality of recording media, such as a CD player adapted to hold a plurality of CDs.

[0129] Although the present invention has been described with reference to a particular preferred embodiment thereof, it is to be understood that numerous modifications and variations can be made in the described approach, without departing from the spirit and scope of the invention.